Landslide susceptibility evaluation method based on machine learning

Abstract

The invention provides a landslide susceptibility evaluation method based on machine learning. The landslide susceptibility evaluation method comprises the following steps: determining an evaluation index and a grading standard; constructing to obtain a landslide susceptibility evaluation model; the landslide susceptibility evaluation model comprises a random forest model, a model hyper-parameter optimization module and a linear weighted combination model; and training the landslide susceptibility evaluation model by using the training sample set to obtain the trained landslide susceptibility evaluation model, and predicting the landslide susceptibility level of each evaluation unit in the research area. The invention provides a landslide susceptibility evaluation method based on machine learning, and the method comprises the steps: accurately obtaining the weight of each evaluation factor through a random forest model, carrying out the linear weighting of each evaluation factor in an evaluation unit, and enabling an obtained comprehensive information amount value to reflect the landslide susceptibility grade of the evaluation unit more accurately, the accuracy of a landslide susceptibility result is improved, and a reliable basis is provided for landslide disaster risk assessment and management.

Description

Technical Field

[0001] The invention belongs to the technical field of landslide disaster risk assessment, and specifically relates to a landslide susceptibility assessment method based on machine learning. Background Art

[0002] Landslide refers to the natural phenomenon that the soil or rock on the slope slides down the slope as a whole or in a dispersed manner along a certain weak surface or weak zone under the action of gravity, affected by factors such as river scouring, groundwater activity, rainwater soaking, earthquakes and artificial slope cutting. Landslides often cause huge losses to industrial and agricultural production and people's lives and property, and sometimes even devastating disasters. Therefore, accurate assessment of the susceptibility of landslides in the study area can provide a basis for landslide disaster risk assessment and management.

[0003] In the prior art, the traditional information volume model is used to directly superimpose various evaluation factors to evaluate the susceptibility of landslides, which makes it difficult to accurately evaluate the susceptibility of landslides. Summary of the invention

[0004] In view of the defects of the prior art, the present invention provides a landslide susceptibility evaluation method based on machine learning, which can effectively solve the above problems.

[0005] The technical solution adopted by the present invention is as follows:

[0006] The present invention provides a landslide susceptibility assessment method based on machine learning, comprising the following steps:

[0007] Step S1, analyzing the key factors affecting the development and distribution of landslides in the study area and determining evaluation indicators; the evaluation indicators include m evaluation factors, expressed as: evaluation indicator f = f 1 ,f 2 ,...,f m ;

[0008] Step S2, determine each evaluation factor f k The grading standard, k = 1, 2, ..., m; the grading standard is used to classify each evaluation factor f k The measured value is converted into the corresponding level;

[0009] Step S3, construct a training sample set:

[0010] Set the evaluation unit size to grid the study area into multiple evaluation units; obtain the evaluation factors f of the evaluation index f of each evaluation unit i at the same historical time k The measured value of each evaluation factor f is obtained according to the grading standard of step S2. k The level of the measured value is obtained, and then the evaluation index level F of the evaluation unit i is obtained.i =(χ i1 ,x i2 ,...,χ im ), where χ i1 ,x i2 ,...,χ im Respectively represent the levels of the measured values ​​of each evaluation factor of the evaluation unit i; further obtain the landslide susceptibility level of each evaluation unit i, and use the landslide susceptibility level of the evaluation unit i as a label to construct a training sample corresponding to the evaluation unit i, which is expressed as: (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il}; Among them, C il represents the landslide susceptibility level of evaluation unit i, which is the training sample label; l = 1, 2, ..., N, N represents the total number of landslide susceptibility levels;

[0011] A number of evaluation units belonging to different landslide susceptibility levels are selected in proportion, that is, a training sample set is constructed;

[0012] Step S4, constructing a landslide susceptibility assessment model; the landslide susceptibility assessment model includes a random forest model, a model hyperparameter optimization module and a linear weighted combination model;

[0013] The landslide susceptibility assessment model is trained using the training sample set to obtain the trained landslide susceptibility assessment model, and the training method is:

[0014] The training samples (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il} Input the random forest model to get each evaluation factor f k The weight P k ;

[0015] The linear weighted combination model shown in formula (1) is used to train the training samples (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il The levels of each evaluation factor in} are linearly weighted to obtain the comprehensive information value y i ;

[0016]

[0017] Compare the comprehensive information value y i and training sample labels C il The deviation is used to obtain the loss; according to the loss, the model hyperparameter optimization module is used to optimize the model hyperparameters of the random forest model until the model converges, thereby obtaining the trained landslide susceptibility assessment model;

[0018] Step S5, using the trained landslide susceptibility assessment model, according to each assessment factor f of each assessment unit in the study area obtained in real time k The measured value is used to predict the landslide susceptibility level of each evaluation unit, so as to obtain the landslide susceptibility level distribution of the study area in real time.

[0019] Preferably, in step S1, the evaluation factors include evaluation factors of terrain factor dimension, evaluation factors of engineering geological rock formation dimension, evaluation factors of hydrological conditions dimension of study area and evaluation factors of intensity dimension of human engineering activities.

[0020] Preferably, the evaluation factors of the terrain factor dimensions include elevation, slope, aspect, plan curvature, profile curvature and NDVI;

[0021] Evaluation factors of the engineering geological rock formation dimensions include topography, rock formation, topographic moisture and faults;

[0022] Evaluation factors of the hydrological condition dimensions of the study area include annual cumulative rainfall, water flow intensity and distance to the river;

[0023] The evaluation factors of the intensity dimension of human engineering activities include distance from residential areas, distance from roads and distance from earthquake sources.

[0024] Preferably, the classification criteria include natural breakpoint method, direction, step size and category.

[0025] Preferably, in step S4, the training sample (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il} Input the random forest model and use formula (2) to get each evaluation factor f k The weight P k :

[0026]

[0027] Where: D Gkhj Represents the kth evaluation factor f kThe reduction value of the Gini index of the jth node in the hth decision tree; h = 1, 2, ..., n, n represents the total number of decision trees included in the random forest model; j = 1, 2, ..., t, t represents the total number of nodes in the hth decision tree; k = 1, 2, ..., m, m represents the total number of categories of evaluation factors;

[0028] Evaluation factor f k The weight P k , represents the evaluation factor f k In all evaluation factors f 1 ,f 2 ,...,f m The importance of .

[0029] Preferably, when each decision tree of the random forest model performs node splitting, an optimal splitting method using impurity measurement is adopted, wherein the impurity is calculated by a Gini index method.

[0030] Preferably, the model hyperparameter optimization module uses a KNN hyperparameter adaptive optimization algorithm to optimize the model hyperparameters of the random forest model.

[0031] The landslide susceptibility assessment method based on machine learning provided by the present invention has the following advantages:

[0032] The present invention provides a landslide susceptibility evaluation method based on machine learning. The random forest model is used to accurately obtain the weight of each evaluation factor, and then each evaluation factor in the evaluation unit is linearly weighted. The obtained comprehensive information value can more accurately reflect the landslide susceptibility level of the evaluation unit, improve the accuracy of the landslide susceptibility result, and further provide a reliable basis for landslide disaster risk assessment and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flow chart of a landslide susceptibility assessment method based on machine learning provided by the present invention;

[0034] Figure 2 A DEM map of the geographical location and altitude of the study area provided in the embodiment of the present invention;

[0035] Figure 3 A schematic diagram of the classification of landslide susceptibility evaluation factors provided by an embodiment of the present invention;

[0036] Figure 4 The objective weight result diagram of the evaluation factors provided by the embodiment of the present invention;

[0037] Figure 5 A landslide susceptibility zoning result map provided by an embodiment of the present invention;

[0038] Figure 6 ROC diagrams for susceptibility assessment of different models provided in embodiments of the present invention. DETAILED DESCRIPTION

[0039] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] The present invention provides a landslide susceptibility evaluation method based on machine learning, which can accurately determine the objective weight of each evaluation factor, reduce the influence of subjective factors on the evaluation factor weight, and improve the accuracy of landslide disaster susceptibility evaluation.

[0041] The present invention provides a landslide susceptibility assessment method based on machine learning, comprising the following steps:

[0042] Step S1, analyzing the key factors affecting the development and distribution of landslides in the study area and determining evaluation indicators; the evaluation indicators include m evaluation factors, expressed as: evaluation indicator f = f 1 ,f 2 ,...,f m ;

[0043] As an embodiment, the evaluation factors include, but are not limited to, evaluation factors of terrain factor dimension, evaluation factors of engineering geological rock formation dimension, evaluation factors of hydrological conditions dimension of the study area, and evaluation factors of intensity dimension of human engineering activities.

[0044] The evaluation factors of the terrain factor dimensions include, but are not limited to, elevation, slope, aspect, plan curvature, profile curvature and NDVI;

[0045] Evaluation factors of the engineering geological rock formation dimensions include, but are not limited to, topography, rock formation, topographic moisture and faults;

[0046] Evaluation factors of the hydrological condition dimensions of the study area, including but not limited to annual cumulative rainfall, water flow intensity and distance from the river;

[0047] The evaluation factors of the intensity dimension of human engineering activities include, but are not limited to, distance from residential areas, distance from roads, and distance from earthquake sources.

[0048] In practical applications, under different research background conditions, the selected evaluation factors have certain differences. In general, engineering geological rock group, slope, elevation and slope aspect are the evaluation factors with the highest selection ratio and the most commonly used basic evaluation factor combination. With the increase in the number of evaluation factors, fault distance, water system, human activities, vegetation, seismic peak acceleration and other evaluation factors are considered in turn. The actual selection of evaluation factors needs to be determined in combination with the previous research foundation and the temporal and spatial distribution law of landslides.

[0049] Step S2, determine each evaluation factor f k The grading standard, k = 1, 2, ..., m; the grading standard is used to classify each evaluation factor f k The measured value is converted into the corresponding level;

[0050] Specifically, combined with the characteristics of each evaluation factor, each evaluation factor f k The grading criteria are the same or different, and the grading criteria include but are not limited to natural breakpoint method, direction, step size and category.

[0051] Step S3, construct a training sample set:

[0052] Set the evaluation unit size to grid the study area into multiple evaluation units; obtain the evaluation factors f of the evaluation index f of each evaluation unit i at the same historical time k The measured value of each evaluation factor f is obtained according to the grading standard of step S2. k The level of the measured value is obtained, and then the evaluation index level F of the evaluation unit i is obtained. i =(χ i1 ,x i2 ,...,χ im ), where χ i1 ,x i2 ,...,χ im Respectively represent the levels of the measured values ​​of each evaluation factor of the evaluation unit i; further obtain the landslide susceptibility level of each evaluation unit i, and use the landslide susceptibility level of the evaluation unit i as a label to construct a training sample corresponding to the evaluation unit i, which is expressed as: (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il}; Among them, C il represents the landslide susceptibility level of evaluation unit i, which is the training sample label; l = 1, 2, ..., N, N represents the total number of landslide susceptibility levels;

[0053] In this step, the landslide susceptibility level can be set to five levels, namely: extremely high susceptibility level, high susceptibility level, medium susceptibility level, low susceptibility level, and extremely low susceptibility level.

[0054] A number of evaluation units belonging to different landslide susceptibility levels are selected in proportion, that is, a training sample set is constructed;

[0055] In practical applications, when constructing a training sample set, the measured values ​​of each evaluation factor of each evaluation unit can be obtained in the following ways:

[0056] The landslide hazard point data comes from the "geological hazard point spatial distribution data" of the Resource and Environmental Science Data Center of the Chinese Academy of Sciences, and the data format is shape file format. The geological lithology data comes from the "China geological lithology spatial distribution data" of the Resource and Environmental Science Data Center of the Chinese Academy of Sciences. The digital elevation model (DEM) data (ASTER-GDEM) with a resolution of 12.5 meters is used to extract topographic and geomorphic attributes and topographic humidity index and water flow intensity index. Meteorological observation data comes from Worldclim to calculate precipitation information. The faults and boundaries of the study area come from the Resource and Environmental Science and Data Center of the Chinese Academy of Sciences to extract evaluation factor parameters. The settlement data comes from the results of the first national geographic survey. Road data and water system data are public data sets provided by OpenStreetMap, etc.

[0057] Step S4, constructing a landslide susceptibility assessment model; the landslide susceptibility assessment model includes a random forest model, a model hyperparameter optimization module and a linear weighted combination model;

[0058] The landslide susceptibility assessment model is trained using the training sample set to obtain the trained landslide susceptibility assessment model, and the training method is:

[0059] The training samples (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il} Input the random forest model to get each evaluation factor f k The weight P k ;

[0060] Specifically, the random forest model is a combined classification model composed of multiple decision trees {h(X,Θk), k=1,2,…}. The parameter set {Θk} is an independent and identically distributed random vector. Given the independent variable X, the optimal classification result is voted by each decision tree model. The present invention uses a random forest model to evaluate the relative weights of various evaluation factors of landslide susceptibility, and these weights are derived based on the Gini index. In the decision tree of the random forest, impurity is used to measure the optimal node splitting method, and the impurity is calculated by the Gini index method. By calculating each evaluation factor f k The reduction in the Gini index when a node splits D Gk , D of all nodes in the random forest Gk After summing, take the average of all trees, which is the evaluation factor f k The relative importance of the evaluation factor is measured by the percentage of its average Gini reduction value to the total average Gini reduction value of all evaluation factors, which can be calculated according to formula (2).

[0061] Specifically, using formula (2), we can get each evaluation factor f k The weight P k :

[0062]

[0063] Where: D Gkhj Represents the kth evaluation factor f k The reduction value of the Gini index of the jth node in the hth decision tree; h = 1, 2, ..., n, n represents the total number of decision trees included in the random forest model; j = 1, 2, ..., t, t represents the total number of nodes in the hth decision tree; k = 1, 2, ..., m, m represents the total number of categories of the evaluation factor; the evaluation factor f k The weight P k , represents the evaluation factor f k In all evaluation factors f 1 ,f 2 ,...,f m The importance of .

[0064] In this step, when each decision tree of the random forest model performs node splitting, the best splitting method using impurity measurement is adopted, wherein the impurity is calculated by the Gini index method.

[0065] Then, the linear weighted combination model shown in formula (1) is used to train the training samples (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C ilThe levels of each evaluation factor in} are linearly weighted to obtain the comprehensive information value y i ;

[0066]

[0067] Compare the comprehensive information value y i and training sample labels C il deviation to obtain the loss; according to the loss, the model hyperparameter optimization module is used to optimize the model hyperparameters of the random forest model. As a preferred embodiment, the model hyperparameter optimization module uses a KNN hyperparameter adaptive optimization algorithm to optimize the model hyperparameters of the random forest model until the model converges to obtain the trained landslide susceptibility assessment model.

[0068] Specifically, since KNN and random forest are two completely different algorithms, the parameter optimization of the KNN model can be achieved through random search, and the random forest training and verification are performed by randomly extracting parameter combinations to find a better parameter configuration. Therefore, the parameter optimization technology of the KNN model is used to initialize the random forest model, and the random forest model is optimized by further adaptive optimization to search for the best parameters.

[0069] Step S5, using the trained landslide susceptibility assessment model, according to each assessment factor f of each assessment unit in the study area obtained in real time k The measured value is used to predict the landslide susceptibility level of each evaluation unit, so as to obtain the landslide susceptibility level distribution of the study area in real time.

[0070] The present invention provides a landslide susceptibility evaluation method based on machine learning. The random forest model is used to accurately obtain the weight of each evaluation factor, and then each evaluation factor in the evaluation unit is linearly weighted. The obtained comprehensive information value can more accurately reflect the landslide susceptibility level of the evaluation unit, improve the accuracy of the landslide susceptibility result, and further provide a reliable basis for landslide disaster risk assessment and management.

[0071] An embodiment is described below:

[0072] Taking Linxia Hui Autonomous Prefecture of Gansu Province as an example, the method proposed in the present invention is used to conduct experiments and compared with the traditional model to verify the feasibility of the method proposed in the present invention.

[0073] The DEM map of the geographical location and altitude of Linxia Hui Autonomous Prefecture, such as Figure 2As shown in Figures (a) and (b) of the 2017 Tibet Autonomous Region, the region is located on the northeastern edge of the Qinghai-Tibet Plateau. Due to the uplift of the Qinghai-Tibet Plateau, the geological structure is very complex and the neotectonic movement is relatively strong, so frequent earthquakes are inevitable. In addition, the environment of the Yellow River Basin is fragile and the landforms are diverse. Thousands of landslides caused by earthquakes pose a serious threat to the lives and property of local residents. According to statistics, most of the landslides in the autonomous prefecture are caused by earthquakes and rainfall.

[0074] With a resolution of 12.5 m × 12.5 m as the evaluation unit size, the study area was divided into 53,131,004 evaluation units, with a total of 542 landslide disaster points. In order to ensure the balance of the samples, 542 random points were randomly selected 1 km away from the disaster points to form the training samples.

[0075] In this embodiment, the traditional information volume model, the traditional weighted information volume model and the weighted information volume model with optimized parameters proposed in the present invention are used to evaluate the susceptibility of landslide disasters in Linxia Hui Autonomous Prefecture.

[0076] Among them: the traditional information volume model is an information volume model that uses statistical methods to calculate the information volume values ​​of different subcategories in each evaluation factor, and uses this to express the degree of its promotion to disaster development. Its formula is:

[0077]

[0078] Where: i Indicates the level of evaluation factor i in the evaluation unit; I(χ i , H) represents the amount of information contributed by the evaluation factor i to the landslide hazard; S is the total number of evaluation units in the study area; S i is the number of evaluation units with evaluation factor i in the study area; N is the total number of evaluation units with disaster points in the study area; N i is the number of evaluation factors i distributed in the landslide disaster area; I is the comprehensive information volume in the evaluation unit; m is the number of evaluation factors.

[0079] Traditional weighted information model: The random forest model of the present invention is used to determine the weight of each evaluation factor, but the KNN hyperparameter adaptive optimization model is not used.

[0080] The parameter-optimized weighted information model: abbreviated as IV-KRF, is the landslide susceptibility evaluation model provided by the present invention, including a random forest model, a model hyperparameter optimization module and a linear weighted combination model.

[0081] like Figure 1 As shown in the figure, the main steps of the study include: (1) selection of evaluation factors; (2) construction of evaluation model; and (3) landslide susceptibility evaluation.

[0082] In terms of the selection of evaluation factors, it is considered that the terrain factor is the air-side condition that controls the slope, and determines the development and distribution of landslides to a large extent. For this reason, elevation, slope, slope aspect, plane curvature, profile curvature and NDVI are extracted as reflection factors of terrain. As the basic constituent unit of the slope, the properties of the rock group have an absolute control effect on the stability of the slope. Therefore, topography, rock group, terrain humidity and faults are selected as evaluation factors. The scouring effect of the river on the embankment, rainwater erosion, and groundwater changes will all lead to the softening of soil texture and cause landslides. Therefore, the annual cumulative rainfall, water flow intensity and distance from the river are selected as evaluation factors reflecting the hydrological conditions of the study area. The areas with frequent human engineering activities in the study area are mostly frequent landslide disaster areas. The distance from residential areas, distance from roads and distance from the earthquake source are selected as evaluation factors reflecting the intensity of human engineering activities.

[0083] The study combines the characteristics of each evaluation factor and adopts a variety of classification standards (natural discontinuity method, direction, step size and category, etc.) to classify the evaluation factors, such as Figure 3 The figure below is a schematic diagram of the classification of landslide susceptibility evaluation factors. Figure 3 a) elevation, b) slope, c) aspect, d) plan curvature, e) profile curvature, f) NDVI, g) topography, h) distance from the epicenter, i) topographic humidity, j) precipitation, k) flow intensity, l) lithology, m) distance from the fault, n) distance from the residential area, o) distance from the river, and p) distance from the road.

[0084] The natural break method is a statistical method for grading and classifying according to the law of numerical statistical distribution, which can maximize the difference between classes. Therefore, rainfall, slope, and elevation are graded according to the natural break method. Slope direction is graded according to direction. Roads and rivers are graded with a step length of 500 meters. Taking into account the extensiveness and complexity of human activities, the distance from residential areas is graded with a step length of 5 kilometers. In view of the integrity and complexity of the rock and soil mass, this embodiment divides the rock group as a whole, and divides the rock group into 8 categories according to the point density of the landslide disaster points in each rock group. Based on the precipitation data, a multi-year average precipitation contour map is generated, and the distribution law and frequency of the precipitation data of the disaster point are statistically analyzed. It is found that the disaster points are mostly located in areas with high precipitation, and the precipitation contour map is rasterized and re-divided into five levels: NDVI and TWI data are obtained using hydrological analysis tools and raster calculators, and then reclassified according to the natural break method.

[0085] Gini index: construct a binary differentiation random forest model. Based on 542 known landslide point data, non-landslide point samples were randomly selected, of which 70% of the sample points were used as training data and 30% of the sample points were used as test data. By adjusting the model hyperparameters, the contribution rate of the evaluation factor (i.e. the relative weight of the evaluation factor) was finally calculated, and a statistical chart was drawn, such as Figure 4 As shown, it is a result diagram of objective weights of evaluation factors, wherein a) is the weight before optimization, and b) is the weight after optimization by the method of the present invention.

[0086] Comparison of susceptibility partition diagrams: Figure 5 As shown in the figure, it is the result map of landslide susceptibility zoning, a) is the information volume model, b) is the weighted information volume model, and c) is the weighted information volume model with parameter optimization. In general, the three models have certain similarities in the susceptibility evaluation results. The information volume model does not take into account the weight of the evaluation factor. Therefore, in the information volume model, there are almost no extremely high susceptibility areas, the high susceptibility areas occupy a large area, and the distinction between low susceptibility areas and medium susceptibility areas is not obvious enough. Moreover, most of the disaster points fall in the low susceptibility areas and below, and it is obvious that the division is not reasonable. Compared with the information volume model, the area of ​​susceptibility areas in the weighted information volume model increases, and the areas of high susceptibility areas and extremely high susceptibility areas decrease. The model pays more attention to medium-risk areas. However, there are still some disaster points scattered freely in the low susceptibility areas and below. In the weighted information volume model with parameter optimization of the present invention, the area of ​​extremely high susceptibility areas has increased significantly, especially in the north and southwest regions, and the zoning is more concentrated and clear. The area of ​​high-prone areas has been reduced, and the areas of medium-prone and low-prone areas have been adjusted accordingly. The density of disaster points is more closely matched with the susceptibility zoning.

[0087] Comparison of susceptibility zoning information statistical tables: As shown in Tables 1, 2 and 3, they are the susceptibility zoning information statistical tables using each model.

[0088] Table 1 Statistics of landslide susceptibility zoning based on information model

[0089]

[0090] Table 2 Statistics of landslide susceptibility zoning based on weighted information model

[0091]

[0092] Table 3 Statistics of landslide susceptibility zoning based on parameter-optimized weighted information model

[0093]

[0094] As can be seen from Table 1, in the evaluation results of the information volume model, the high-risk area occupies the largest area proportion (70.17%), followed by the medium-risk area (22.23%). The high-risk area has the highest proportion of disaster points (86.90%), with a density of 0.0822 / km -2 The density of disaster points in the extremely high-risk area is the highest, at 0.1712 / km -2 .

[0095] As can be seen from Table 2, in the evaluation results of the weighted information model, the area proportion of the high-prone area decreased to 59.07%, and the medium-prone area increased to 30.53%. The proportion of disaster points in the high-prone area decreased to 71.59%, but the density of disaster points in the extremely high-prone area further increased to 0.1960 / km -2 .

[0096] As can be seen from Table 3, in the evaluation results of the weighted information volume model IV-KRF with parameter optimization of the present invention, the area proportion of the high-risk area increased again to 70.37%, and the medium-risk area decreased significantly to 14.61%. The proportion of disaster points in the extremely high-risk area increased significantly to 22.69%, and the density decreased slightly to 0.1824 / km -2 .

[0097] By comparing the three models, it can be seen that in terms of area distribution: the high-prone area always occupies a dominant position, but the weighted information volume model and the weighted information volume model with optimized parameters have some adjustments in the area distribution of the medium-prone area and the extremely high-prone area. In terms of disaster point density: the extremely high-prone area shows a higher density of disaster points in the weighted information volume model, and the weighted information volume model with optimized parameters has some balance in this regard, but still maintains a high density.

[0098] Precision verification: As shown in Table 4, it is a statistical table of the precision of non-landslide sample selection methods for different models:

[0099] Table 4 Accuracy statistics of different non-landslide sample selection methods

[0100] Sample selection method AUC value Standard error Information model (a) 0.738 0.045 Weighted information model (b) 0.775 0.089 IV-KRF(c) 0.854 0.062

[0101] It can be seen from Table 4 that in the parameter-optimized weighted information model of the present invention, the AUC value is 0.854, indicating that the model has a good prediction effect and high reliability.

[0102] The ROC curve is a method commonly used to evaluate model accuracy. It reveals the relationship between sensitivity and specificity through the comprehensive index of two continuous variables. At present, most studies on geological hazard susceptibility assessment use the ROC curve method to verify the accuracy of the assessment results.

[0103] The ROC curve is a commonly used detection method. Its abscissa represents 1 - specificity, and its ordinate represents sensitivity. The numerical range of the ROC curve is from 0 to 1. The steeper the curve, the higher the prediction accuracy of the model. The quality of the model can be judged by the value of the area under the ROC curve (AUC). The higher the AUC value, the stronger the prediction ability of the trained model. For example, AUC > 0.9 indicates a very good prediction effect; 0.8 < AUC ≤ 0.9 indicates a good prediction effect; 0.5 < AUC ≤ 0.8 indicates an average prediction effect; AUC equal to 0.5 means the model result has no reference value; and AUC less than 0.5 indicates that the model does not conform to the actual situation.

[0104] As Figure 6 shown, it is the ROC diagram for the susceptibility assessment of different models; among them, a) is the information model, b) is the weighted information model, and c) is the weighted information model with parameter optimization, IV - KRF.

[0105] From Figure 6 it can be seen that the AUC values of the information model, the weighted information model, and the weighted information model with parameter optimization are 0.738, 0.775, and 0.854 respectively. The analysis results show that the accuracy of using the model proposed by the present invention for landslide hazard susceptibility assessment is better than that of other models.

[0106] The above - mentioned is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art of this technology, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A landslide susceptibility assessment method based on machine learning, characterized in that: The following steps are involved: Step S1, analyzing the key factors affecting the development and distribution of landslides in the study area and determining evaluation indicators; the evaluation indicators include m evaluation factors, expressed as: evaluation index f = f1, f2, ..., f m ; Step S2, determine each evaluation factor f k The grading standard, k = 1, 2, ..., m; the grading standard is used to classify each evaluation factor f k The measured value is converted into the corresponding level; Step S3, construct a training sample set: Setting the evaluation unit size to grid the study area into multiple evaluation units; At the same time in history, obtain the evaluation factors f of the evaluation index f of each evaluation unit i k The measured value of each evaluation factor f is obtained according to the grading standard of step S2. k The level of the measured value is obtained, and then the evaluation index level F of the evaluation unit i is obtained. i =(χ i1 ,x i2 ,...,χ im ), where χ i1 ,x i2 ,...,χ im Respectively represent the levels of the measured values ​​of each evaluation factor of the evaluation unit i; further obtain the landslide susceptibility level of each evaluation unit i, and use the landslide susceptibility level of the evaluation unit i as a label to construct a training sample corresponding to the evaluation unit i, which is expressed as: (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il }; Among them, C il represents the landslide susceptibility level of evaluation unit i, which is the training sample label; l = 1, 2, ..., N, N represents the total number of landslide susceptibility levels; A number of evaluation units belonging to different landslide susceptibility levels are selected in proportion, that is, a training sample set is constructed; Step S4, constructing a landslide susceptibility assessment model; the landslide susceptibility assessment model includes a random forest model, a model hyperparameter optimization module and a linear weighted combination model; The landslide susceptibility assessment model is trained using the training sample set to obtain the trained landslide susceptibility assessment model, and the training method is: The training samples (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il } Input the random forest model to get each evaluation factor f k The weight P k ; The linear weighted combination model shown in formula (1) is used to train the training samples (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il The levels of each evaluation factor in} are linearly weighted to obtain the comprehensive information value y i ; Compare the comprehensive information value y i and training sample labels C il The deviation is used to obtain the loss; according to the loss, the model hyperparameter optimization module is used to optimize the model hyperparameters of the random forest model until the model converges, thereby obtaining the trained landslide susceptibility assessment model; Step S5, using the trained landslide susceptibility assessment model, according to each assessment factor f of each assessment unit in the study area obtained in real time k The measured value is used to predict the landslide susceptibility level of each evaluation unit, so as to obtain the landslide susceptibility level distribution of the study area in real time.

2. The landslide susceptibility assessment method based on machine learning according to claim 1, characterized in that: In step S1, the evaluation factors include evaluation factors of terrain factor dimension, evaluation factors of engineering geological rock formation dimension, evaluation factors of hydrological conditions dimension of study area and evaluation factors of intensity dimension of human engineering activities.

3. The landslide susceptibility assessment method based on machine learning according to claim 2 is characterized in that: The evaluation factors of the terrain factor dimensions include elevation, slope, aspect, plan curvature, profile curvature and NDVI; Evaluation factors of the engineering geological rock formation dimensions include topography, rock formation, topographic moisture and faults; Evaluation factors of the hydrological condition dimensions of the study area include annual cumulative rainfall, water flow intensity and distance to the river; The evaluation factors of the intensity dimension of human engineering activities include distance from residential areas, distance from roads and distance from earthquake sources.

4. The landslide susceptibility assessment method based on machine learning according to claim 1, characterized in that: The classification criteria include natural breakpoint method, direction, step size and category.

5. The method for evaluating landslide susceptibility based on machine learning according to claim 1, characterized in that: In step S4, the training sample (F i ,C il )={(χ i1 ,x i2 ,...,χ im ),C il } Input the random forest model and use formula (2) to get each evaluation factor f k The weight P k : Where: D Gkhj Represents the kth evaluation factor f k The reduction value of the Gini index of the jth node in the hth decision tree; h = 1, 2, ..., n, n represents the total number of decision trees included in the random forest model; j = 1, 2, ..., t, t represents the total number of nodes in the hth decision tree; k = 1, 2, ..., m, m represents the total number of categories of evaluation factors; Evaluation factor f k The weight P k , represents the evaluation factor f k Among all evaluation factors f1,f2,...,f m The importance of .

6. The method for evaluating landslide susceptibility based on machine learning according to claim 5, characterized in that: When each decision tree of the random forest model performs node splitting, an optimal splitting method using impurity measurement is adopted, wherein the impurity is calculated by a Gini index method.

7. The method for evaluating landslide susceptibility based on machine learning according to claim 1, characterized in that: The model hyperparameter optimization module uses the KNN hyperparameter adaptive optimization algorithm to optimize the model hyperparameters of the random forest model.

Images