Evaluation and analysis of statistical and coupling models for highway landslide susceptibility

Abstract Landslides have a great impact on the normal traffic of highway, and maintaining the normal traffic of highway is the foundation of economic development, so landslide susceptibility mapping is very important. In this study, four counties, which locate in the central Ganzi Tibetan Autonomous Prefecture, Sichuan Province, China, are taken as the research region. Based on the 190 historical landslide disaster points in the region, six factors-elevation, slope, aspect, plan curvature, profile curvature and TWI (Topographic Wetness Index) - are finally selected for calculation. A landslide disaster is evaluated by two single models of CF (Certainty Factors) and IV (Information Value) models and four coupling models of CF-AHP (Analytic Hierarchy Process), CF-LR (Logistic Regression), IV-AHP and IV-LR models. The accuracy of the six models is evaluated by the ROC (Receiver Operating Characteristic) curve and the Sridevi Jadi parameters. The IV-AHP model has the highest value of 0.9189, which indicates that the IV-AHP model is more appropriate for landslide disaster assessment in the whole region. In the Sridevi Jadi parameters, the IV model have the highest value of 0.8696, showing that the IV model have the highest accuracy in landslide susceptibility assessment in high- and very high-susceptibility regions.


Introduction
A landslide (Chen et al. 2022) refers to the natural phenomenon that the soil or rock on the slope slides partially or completely downward under the action of surface water scouring, soaking or earthquake and human activities.A landslide is one of the most common natural disasters.According to the 2020 Global Natural Disaster Assessment Report, there were 19 landslides all over the world, which affected 179800 people and caused a direct economic loss of $130 million (Beijing Normal University 2021).Moreover, China is one of the countries most affected by landslides.On the basis of the bulletin of the Ministry of Natural Resources of the People's Republic of China (2018China ( -2022) ) in 2020, there were 7480 geological disasters in China, with a direct economic loss of CNY 5.02 billion, of which landslides accounted for 61.4% of the total number of geological disasters, which shows the great losses caused by landslides.Therefore, the landslide susceptibility mapping is essential.
The landslide disaster evaluation model has been continuously developed and improved.Most of the original landslide disaster evaluation models are statistical methods.Joel et al. (2021) used the FR (Frequency Ratio) model to assess the vulnerability of landslide in the Cross River State of Nigeria, and the results of which will be useful for landslide vulnerability prevention and land use planning in the research region, and provide experience for other regions with similar geographical environment.Jin et al. (2022) applied LR (Logistic Regression) model to assess the sensibility of landslide in the Bailong River Basin of China.The results are helpful to the early identification and disaster assessment of landslide disasters in Bailong River Basin and similar alpine areas.Chen et al. (2021) used the WOE (Weight Of Evidence) model to evaluate the landslide sensitivity of the highways in Hubei section of the Three Gorges Reservoir area, and established the comprehensive index system and method of highway landslide sensitivity evaluation, which provides a reference for risk evaluation and prevention management.Zhong (2015) applied the IV (Information Value) model to analyse the stability of bank slopes in three parallel river region of the Sichuan-Tibet Railway and established a personal geographic database of adverse geological phenomena in the research region to lay foundation for geological information of railroad alignment selection.Fan et al. (2017) coupled the CF (Certain Factor) and AHP (Analytic Hierarchy Process) to assess the susceptibility of landslide disaster in Ziyang area of the Qinba Mountain in China, and the results have some reliability.Chen et al. (2016) used IV and LR models to map the landslide sensitivity in Zigui-Badong area near the Three Gorges reservoir, and the results show that both methods are feasible for the landslide sensitivity mapping in the research region.Alsabhan et al. (2022) used WOE, IV and FR models to evaluate the landslide sensitivity near the Himalayas, and the landslide sensitivity zoning maps obtained can reduce the economic and human losses associated with landslides in highly sensitive and extremely sensitive areas.
With the development of RS and GIS, machine learning is gradually applied to landslide disaster assessment.Pham et al. (2021) applied FDEMATEL (Fuzzy Decision Making Trial and Evaluation Laboratory)-ANP (Analytic Network Process), FR, LR, RFC (Random Forest Classifier), NBC (Naïve Bayes Classifier) and XGBoost (Extreme Gradient Boosting) to assess the landslide susceptibility in the Kysuca river basin, Slovakia, the result show that RFC is the most accurate model.Costache et al. (2021) used CF, FAHP (Fuzzy AHP), XGBoost-CF and DLNN (Deep Learning Neural Network)-CF models to evaluate the flood-induced landslide susceptibility in the Indian state of Assam, and the results reveal that XGBoost-CF is the most performance model.El-Magd et al. (2021) applied RF (Random Forest), KNN (K-Nearest Neighbor) and NB (Naïve Bayes) to predict the landslides events in Jabal Farasan, Jeddah, Saudi Arabia, the results show that the KNN model is most appropriate.Ali et al. (2020) used FDEMATEL-ANP, NB and RF model to predict the landslide susceptibility in the Kysuca river basin, Slovakia, the results show that the RF is the optimum model.The advantages of machine learning are significant; it has strong learning ability, and can obtain the analysis results without too much prior knowledge; it can better reflect the complex and nonlinear relationship between susceptibility levels and influence factors, and has good data processing ability.However, there will be over-fitting phenomenon when the amount of calculation is large.Therefore, for the areas with frequent landslides, and the prediction accuracy cannot be guaranteed only by using a single model for landslide susceptibility evaluation.Especially in the areas with complex geological structures, the coupling and comparison of different methods can improve the accuracy of landslide susceptibility evaluation and ensure the stability of model prediction in areas with complex geological structures.
Previous studies have shown that susceptibility (United Nations Office for Disaster Risk Reduction 2022) is a 'process, phenomenon or human activity that may cause loss of life, injury or other health impacts, property damage, social and economic disruption or environmental degradation'.The origins of susceptibility can be natural or human-made.Vulnerability (Ferreira et al. 2021) describes 'the conditions determined by physical, social, economic and environmental factors or processes which increase the susceptibility of an individual, a community, assets or systems to the impacts of hazards'.Risk (Ni et al. 2010) is considered to be the probability of a loss, and affects by susceptibility, vulnerability and so on.Therefore, this article focus on the landslide susceptibility mapping.
Most of the previous studies have focused on studying landslides caused by earthquake, mountains and heavy precipitation, but in fact the damage caused by landslides to highway cannot be ignored.Because of its characteristics, highways are bound to be affected by landslides.At the same time, highway construction may also cause landslides, which often lead to highway traffic being blocked or interrupted, endangering pedestrians' travel safety, causing people's lives and property losses, increasing the costs of maintenance and governance, and becoming adverse factors affecting social life order and restricting regional economic exchanges and economic development.Therefore, facing the demand of highway landslide prevention and mitigation, it is of great significance to evaluate the susceptibility of highway landslide on the basis of studying the influencing factors of highway landslide.To fill this gap, in this article, 190 highway damage points caused by landslides were collected, and nine evaluation factors-elevation, slope, aspect, curvature, plan curvature, profile curvature, height difference, TWI (Topographic Wetness Index) and surface roughness-are preliminarily selected, and the independence of each evaluation factors is tested by using the correlation coefficient, finally, a total of six evaluation factors-elevation, slope, aspect, plan curvature, profile curvature and TWI-are selected to assess the susceptibility of landslide in the central Ganzi Tibetan Autonomous Prefecture by using CF, IV, CF-AHP, CF-LR, IV-AHP and IV-LR models.The landslide susceptibility map obtained not only helps the highway infrastructure maintenance business to carry out accurately, but also plays an important role in supporting data for timely rescue, post-disaster reconstruction and other work after the occurrence of landslides.

Research region
The research region is in the central Ganzi Tibetan Autonomous Prefecture, Sichuan Province, China (as shown in Figure 1(a)), and it contains four counties, namely, Yajiang, Daofu, Xinlong and Litang, between 28 27 0 -31 41 0 and 99 24 0 -101 44 0 , with the whole area of 36858 km 2 and the highway is relatively dense.The terrain in the research region belongs to the transition zone between Yunnan-Guizhou Plateau and Sichuan Basin, and its high in the west and low in the east, and the central and western regions are river valleys.Its climate mainly belongs to the Qinghai-Tibet Plateau, which shows an obvious vertical distribution with the change of elevation, and it is characterized by low temperature, long winter, little precipitation and enough sunshine.The 190 landslide disaster points shown in this article are manually collected data in the field, which are highway damage points that caused damage after previous landslides, and are plotted by collecting information such as latitude and longitude, disaster type, geographic location, and highway section name of the highway damage points.The 190 disaster points are mainly distributed in the northwest and east (as shown in Figure 1(b)), where the population density is high and human activities are more frequent.

Data sources and uses
The data sources of this research are as follows: (1) Landslide disaster points data, which is used to analyse the distribution of landslide points; (2) DEM data with 30 m resolution, which is used to extract data such as elevation, height difference and curvature of the research region; (3) The 90 m resolution data of STRM-SLOPE and STRM-ASPECT are used to extract the slope and aspect of the research region.The data sources and uses are shown in Table 1.

Research technique
The main research technique flow diagram is shown in Figure 2.

CF
CF was first proposed by Shortliffe and Buchanan (1975) in 1975, and further improved by Heckerman (1986) in 1986.CF (Qin et al. 2021;Song et al. 2021) is a possibility function applied to analyse the influence degree of each factor that affecting the occurrence of an event, and it considers that the landslide susceptibility can be determined depend on the statistical relationship between previous disaster and environmental factors, and its formula is as follows (Ali et al. 2022): where PPa is the conditional probability of a certain type of event in the data a, which is expressed in the landslide susceptibility assessment as the ratio of the amount of disaster points of a certain type of event in the factor a to the area of the corresponding event; PPs is the prior probability of occurrence of a certain type of event in the whole research region, which can be expressed as the ratio of the amount of disaster points in the whole research region to the area of the research region in the landslide susceptibility assessment; CF is the certainty coefficient, with the value range from À1 to 1.The larger the value of CF, the higher the certainty of the event occurrence, that is, the higher the susceptibility of landslide disaster in the corresponding area.

IV
IV (Tang et al. 2020;Tang et al. 2021) is a statistical analysis method, which is widely used in landslide susceptibility assessment.It is assumed that some variables have an influence on the occurrence of disaster, and the contribution of these variables to the occurrence of disaster is different, then the degree of their impact can be expressed quantitatively through the amount of information.The amount of information can be defined by the following formula: I Y, x 1 , x 2 , :::, where I is the amount of information of the combination of variables x 1 , x 2 , :::, x n , PðY, x 1 , x 2 , :::, x n Þ is the joint probability of landslide event under the variable x i , and PðYÞ is the possibility of event occurrence.In practical applications, the probability calculation is not convenient, so the area ratio can be used instead.The formula is as follows: where N i is the amount of landslide points distributed under variables x i in the research region, N is the whole amount of landslide points in the research region, S i is the region containing variables x i in the research region, and S is the total area of the research region.

AHP
First proposed by Staay in the early 1970s, this is a simple, flexible and practical multi-criteria decision method and has been widely used in landslide susceptibility assessment (Feizizadeh et al. 2014;Wang and Li 2017).AHP (Xu et al. 2018;Xu 2020) first divides the decision-making problem into decision-making objectives, middle-level elements and alternatives.Second, each factor is compared pairwise and a judgment matrix is established.Next, the hierarchical single sorting and consistency test are carried out, that is, the factors are weighted.Finally, the overall ranking of levels and consistency test are carried out.After all requirements are met, the weight can be got.The consistency test calculation formula is as follows: where, CR is its consistency ratio, CI is the consistency index of the matrix, RI is its random consistency index, whose value depends on its order, as shown in where P is the possibility of landslide occurrence (range 0-1), Z is the sum of all control variables based on weight, X n ðn ¼ 1, 2, :::, nÞ is independent variable, b n ðn ¼ 0, 1, :::, nÞ is the regression coefficient based on the training samples.
2.3.5.Accuracy analysis methods 1. ROC (Receiver Operating Characteristic) curve A ROC curve (Wang 2020) is the most commonly used accuracy analysis method in landslide susceptibility assessment.The abscissa of the ROC curve represents the false positive rate (specificity), which indicates the cumulative percentage of the region of the landslide disaster level from high to low; the ordinate represents the true positive rate (sensitivity), which indicates the cumulative percentage of the amount of disaster points in the corresponding landslide susceptibility level.In order to represent the assessment results more clearly, the size of AUC (Area Under the Curve) (Basharat et al. 2016) is usually applied to indicate the accuracy of the model, and the size of AUC varies from 0 to 1.Meanwhile, the larger the value is, the higher the accuracy and the stronger the applicability of the model.

Sridevi Jadi parameter
The accuracy assessment method proposed by Sridevi Jadi (Qiu et al. 2014) is expressed in empirical probability form as: where P is the precision size, K S is the amount of landslide points in the region with high and very high landslide susceptibility level in the research region, S is the whole amount of landslide points in the research region, K is the area with high and very high landslide susceptibility in the research region, and N is the total area of the research region.

Test the independence of evaluation factors
In order to ensure that the selected evaluation factors meet the requirements of model analysis, the independence of each factors needs to be tested before correlation calculation, so as to eliminate the factors with large correlation to guarantee the accuracy of the model (Tian et al. 2016).The correlation coefficient between each factor is shown in Table 3: Based on the data of Table 2, the correlation coefficient between curvature, height difference, surface roughness and a factor is greater than 0.3 and needs to be eliminated, so a total of six factors, namely elevation, slope, aspect, plan curvature, profile curvature and TWI, can be included in the model for calculation.

Calculation of CF and IV values of evaluation factors
Most of the classification numbers and intervals of evaluation factors are determined by experience, and this will seriously affect the evaluation results of landslides.Therefore, in this article, the classification numbers and intervals are determined by inflection point method and natural break method, respectively (Che et al. 2020;Li and Xu 2020), which classifies elevation, slope, aspect, plan curvature, profile curvature and TWI into 5, 5, 9, 3, 3 and 5 categories respectively.

Elevation
The elevation of the research region varies from 2209 to 6123 m (as shown in Figure 3 (a)).Based on the data of Table 4, the disaster points are most distributed in the region with an elevation between 2209 and 3313 m, and the CF and IV values in this region are also the highest, indicating that the landslide susceptibility in this region is the highest, while the amount of disaster points is the lowest in the region with elevations ranges from 4087 to 4373 m and from 4373 to 6123 m, the values of CF and IV are the lowest in the region with elevation ranges from 4373 to 6123 m, and the landslide susceptibility in this region is the lowest.

Slope
The slope of the research region varies from 0 to 72.517 (as shown in Figure 3 (b)).Based on the data of Table 4, the disaster points are most distributed in the region with slope between 25.594 and 32.988, and the CF and IV values in this region are the highest, indicating the landslide susceptibility is the highest; disaster points are evenly distributed in the other four regions, but the values of CF and IV in the region with slope between 9.669 and 17.916 are the lowest, indicating the susceptibility of landslide disaster is the lowest.

Aspect
The aspect in the research region varies from À 1 to 360 (as shown in Figure 3 (c)).
Based on the data of Table 4, the disaster points are least distributed in the gentle region with the aspect between À 1 and 0, and evenly distributed in other regions, but the values of CF and IV are the highest in the region with the slope between 112.5 and 157.5, indicating that the landslide susceptibility in this region is the highest; the region with the aspect between 292.5 and 337.5 is the lowest, indicating that the landslide susceptibility in this area is the lowest.

Plan curvature
The plan curvature of the research region varies from À0.876 to 0.852 (as shown in Figure 3 (d)).Based on the data of Table 4, the disaster points are most distributed in the region with plan curvature between À0.05 and 0.05, and the other two regions are evenly distributed, but the values of CF and IV are the highest in the region with plan curvature between À0.876 and À0.05 and the lowest in the region with plan curvature between 0.05 and 0.852.

Profile curvature
The profile curvature of the research region varies from À1.048 to 1.308 (as shown in Figure 3 (e)).Based on the data of Table 4, the disaster points are least distributed in

Landslide susceptibility mapping
In this article, two single models (CF, IV) and four coupling models (CF-AHP, CF-LR, IV-AHP, IV-LR) are applied to evaluate the landslide susceptibility in the research region, and make use of the natural break method to divide the result map into five categories: very low, low, moderate, high and very high (as shown in Figure 4 and Table 5).According to Figure 4, the very high-susceptibility region is mainly concentrated in the regions with low elevation and high slope, where human activities are frequent and are more likely to cause landslide disasters.Meanwhile, the very low-susceptibility region is mainly concentrated in the regions with high elevation and low slope, where the influence of human activities is smaller.And among the four counties, Litang County's very high-susceptibility region is relatively small, while Yajiang County's very high-susceptibility region is relatively large.

CF and IV models
The CF and IV values of various evaluation factors in the research region are overlaid by ArcGIS to gain the CF and IV models' result maps of landslide susceptibility in the research region, respectively (as shown in Figure 4 (a) and (b)).

CF-AHP coupling model
The importance level of each factor is determined based on the results of existing studies and a judgment matrix is established (Shown in Table 6).
Once the weight of all factors is gained, the CF value of each factor is weighted and overlaid by using the superposition function of ArcGIS to calculate the result of CF-AHP model of landslide susceptibility in the research region (as shown in Figure 4 (c)).

CF-LR coupling model
In total, 380 sample points are used as the original data of LR model, including 190 landslide disaster points and equivalent non-landslide disaster points which generate by ArcGIS.The CF value of each factor is taken as the independent variable and whether or not there is a landslide as the dependent variable.Logistic regression analysis is carried out based on SPSS software to gain the regression coefficients of each independent variable and take them into equation ( 8): where X 1 is elevation, X 2 is slope, X 3 is aspect, X 4 is plan curvature, X 5 is profile curvature and X 6 is TWI.The coefficient of each factor gained is the factor weight, and the CF value of each factor is weighted and overlaid by using the superposition function of ArcGIS to gain the CF-LR model of landslide susceptibility in the research region (as shown in Figure 4 (d)).

IV-AHP coupling model
According to the weight of each factor, the IV value of each factor is weighted and overlaid by using the superposition function of ArcGIS to gain the result map of IV-AHP model of landslide susceptibility in the research region (as shown in Figure 4 (E)).

IV-LR coupling model
Based on the data of 190 existing landslide disaster points and equivalent non-landslide disaster points in the research region, logistic regression analysis is carried out based on SPSS software.The IV value of each factor is taken as an independent variable to gain its regression coefficients, and the results are brought into Equation ( 9): where X 1 is elevation, X 2 is slope, X 3 is aspect, X 4 is plan curvature, X 5 is profile curvature and X 6 is TWI.
The coefficient of each factor obtained is the factor weight.The IV value of each factor is weighted and overlaid by using the superposition function of ArcGIS to gain the result map of IV-LR model of landslide susceptibility in the research region (as shown in Figure 4(f)).

The distribution of landslides
The distribution of landslides is the most basic analysis of landslide susceptibility assessment results.It can roughly determine the distribution of each models' susceptibility grades, and then initially judge the reliability of the models.Comparison of different susceptibility grades of six models, percentages of different susceptibility grades of six models and comparison of disaster point density of six models are showing in Figure 5, 6 and 7.
According to Figure 5, the width of the extended branch corresponds to the size of the proportion of the data.In the CF, CF-AHP and CF-LR models, the lowsusceptibility region had the highest proportion and the very high-susceptibility region had the lowest proportion.In the IV, IV-AHP and IV-LR models, the very low-susceptibility region had the highest proportion and the very high-susceptibility region had the lowest proportion.
According to Figure 6, the six models -CF, IV, CF-AHP, CF-LR, IV-AHP and IV-LR modelare all divided into five grades.In the very low-susceptibility region, the CF-AHP model has the lowest percentage and the IV-AHP model has the highest percentage.In the low-susceptibility region, the IV-LR model has the lowest percentage and the CF-AHP model has the highest percentage.In the moderate-risk region, the CF-AHP model has the lowest percentage and the CF-LR model has the highest percentage.In the high-risk region, the CF-AHP model has the lowest percentage and the CF model has the highest percentage.In the very high-susceptibility region, the CF-AHP model has the lowest percentage and the IV model has the highest percentage.Among the percentages occupied by the different grades, compared with the transitional regions such as the low-, moderate-and high-susceptibility regions, the very high-and very low-susceptibility regions have a greater impact on the model accuracy.

Accuracy analysis
For testing the accuracy of six landslide disaster models, a ROC curve and Sridevi Jadi parameter are used to verify the results.
In the ROC curve, the accuracy of the single CF model and IV model is 0.9033 and 0.9180, respectively, and the accuracy of the coupling CF-AHP, CF-LR, IV-AHP and IV-LR models are 0.9149, 0.9174, 0.9189 and 0.9172, respectively (as shown in Figure 8).The accuracy of all six assessment models is greater than 0.7500, which shows that all these models have certain reliability.Among them, the single CF model has the lowest accuracy and the coupling IV-AHP model has the highest accuracy, which shows that the IV-AHP model has the best results in the overall landslide susceptibility assessment in the research region.
In the assessment results of the Sridevi Jadi parameters, the accuracy of the single CF model and IV model is 0.7988 and 0.8696, respectively, and the accuracy of the coupling CF-AHP, CF-LR, IV-AHP and IV-LR models are 0.8422, 0.8688, 0.8548 and 0.8655, respectively.Among the six assessment models, the single CF model has the lowest accuracy, and the single IV model have the highest accuracy, which shows that the IV model is the most suitable for landslide susceptibility assessment in high-and very-high susceptibility regions in the research region.

The comparison of these models
In this article, six models-CF, IV, CF-AHP, CF-LR, IV-AHP, IV-LR-are applied to assess landslide susceptibility in the research region.The CF model can determine regional landslide susceptibility based on the relationship between the past disaster points and disaster-causing factors, but it cannot reflect the contribution of each factor to the landslide susceptibility.The IV model builds an assessment and prediction model based on the calculated amount of information, which has more advantageous in the research region with a large number of units, but it cannot reflect the difference in the influence degree of various disaster-causing factors, and there may be a factor that has a suppressive effect on the occurrence of landslide disaster, which is not consistent with the evaluation model.The AHP can gain the weight of each disaster-causing factor by establishing a judgment matrix, and the importance of each factor is reflected in CF-AHP and IV-AHP models, but if there are too many factors and a strong correlation between the factors, the model will be distorted.The LR operation process is not affected by subjective factors, and the physical meaning of the evaluation results is clear, which makes up for the shortcomings of the CF and IV models to a certain extent, but it requires a large number of sample statistics to support, and the reliability of the results will be reduced if the sample data are insufficient.
It is clear from the above analysis that different methods have certain limitations and scope of application.When studying the landslide susceptibility in a certain area, it is necessary to make full use of various data and existing disaster points information, and compare the advantages and disadvantages of various models, so as to select the most suitable model.

Novelty of the application
The National Comprehensive Natural Disaster Risk Census highway carrier census began in 2020, a fundamental effort to improve natural disaster prevention and control capabilities.Along with the rapid development of China's transportation infrastructure construction and the increasing frequency and intensity of natural disasters, the losses and social costs caused by transportation disruptions far exceed the scope and losses of infrastructure damage.Therefore, it is extremely important to find out the status of highway disasters and their distribution, and to do a good job of statistical analysis of highway disaster damage data.The 190 disaster damage points used in this article are all highway damage points caused by landslides collected in the field, and the landslide susceptibility map generated is also a rank map of the hazards caused by landslides to highway.The landslide susceptibility maps obtained not only help the accurate development of road infrastructure maintenance operations, but also play an important role in the supporting data for timely rescue and post-disaster reconstruction after landslides occur.

Reasons why a certain factor has a higher influence
The analysis of the six models in this article found that the elevation is the most influential among the six factors, followed by the slope.This is mainly because different elevation will lead to different precipitation, temperature, vegetation and other factors, and the topography of the research region is complex, and the relative height difference can be as high as 4000 m, which also creates conditions for landslides to occur.According to Table 4, it can be seen that both CF and IV values are maximum at the elevation from 2161 to 3312 m, where human activities are frequent, which also increases the damage of the highway.And as the slope increases, the shear force of the accumulation on the slope also increases, and the stability of the slope decreases, which increases the possibility of landslide occurrence.

Managerial and policy implications
Landslide comprehensive management is a complex and systematic project.Firstly, according to the process of disaster prevention, preparation, response and recovery, a landslide susceptibility management system should be established.Secondly, set up a full-time landslide disaster reduction agency and improve the coordination mechanism.Finally, it is necessary to strengthen the ability of landslide disaster prevention and improve the disaster prevention level of the group.
In terms of policy, firstly, land development in very high-susceptibility regions can be restricted by land planning and land approval; secondly, the government can requisition land and change its use; thirdly, it can prevent risks by improving architectural design and slope stabilization measures.

Summary and key findings
Taking the four counties in the central Ganzi Tibetan Autonomous Prefecture of Sichuan Province as the research region, this article selects six evaluation factors, and uses two single models of CF and IV model and four coupling models of CF-AHP, CF-LR, IV-AHP and IV-LR models to evaluate the landslide susceptibility, and the accuracy of six models is evaluated by ROC curve and Sridevi Jadi parameter.The results can provides experience and reference for landslide susceptibility assessment in other regions.

Selection of evaluation factors
The evaluation factors for landslide susceptibility mapping involve many elements such as humanities, climate, vegetation, soil, water bodies, etc.The evaluation factors in this article are based on previous studies, summarized a large amount of domestic and foreign related literature and combined with the actual situation of the research region, and there is not yet a set of evaluation factors suitable for all regions.In the future, we will further strengthen the use of remote sensing data to extract more evaluation factors and find a set of universal evaluation factors.

Determination of the evaluation factor weight assignment scheme
In this article, AHP is a combination of different experts' opinions to obtain the weights of each factor, while LR is a logistic regression coefficient of each factor obtained from sample data.Both of these weighting schemes have certain deviations, and in future research, the factor assignment schemes will be further improved by combining field surveys and objective assignment methods.

Removing the correlation of different evaluation factors
Landslide susceptibility is a complex process of continuous change influenced by many factors, it is important to determine the dominant factors affecting landslide susceptibility in a scientific and reasonable way.In addition, each evaluation factor is not directly involved in the process of landslide susceptibility, so it is also important to remove the correlation between different factors.Therefore, studying the dominant factors affecting landslide susceptibility and removing the correlation between different factors is still one of the challenges facing landslide susceptibility mapping at present.

Figure 1 .
Figure 1.Overview of the research region.

Figure 2 .
Figure 2. The main research technique flow diagram.

Figure 4 .
Figure 4. Evaluation results of six models.

Figure 6 .
Figure 6.Percentages of different susceptibility grades of six models.

Figure 5 .
Figure 5.Comparison of different susceptibility grades of six models.

Figure 7 .
Figure 7.Comparison of disaster point density of six models.

Table 1 .
Data sources and uses.

Table 2
variables are multiple influencing factors and the dependent variables are whether it is a landslide.The relevant formula is as follows:

Table 2 .
Matrix order and RI value.

Table 3 .
Correlation coefficient matrix of each factor.

Table 4 .
Grading of evaluation factors and calculation the values of CF and IV.the region with profile curvature between À1.048 and À0.05, and the CF and IV values in this region are also the lowest, with the lowest susceptibility of landslide disaster; the disaster points are evenly distributed in the other two regions, but the CF and IV values are the highest in the region with profile curvature between 0.05 and 1.308, which has the highest landslide susceptibility.Based on the data of Table4, the disaster points are least distributed in the region with TWI between 12.136 and 22.710, but the CF and IV values in this region are the highest, and the landslide susceptibility is the highest; there is little difference in the distribution of disaster points in the other four regions, but the values of CF and IV are the lowest in the region with TWI between 9.235 and 12.136, and the landslide susceptibility in this region is the lowest.

Table 5 .
Calculation of evaluation results of six models.

Table 6 .
Judgment matrix and weight of each factor.
Laboratory of Geo-Information Engineering in Surveying, Mapping and Remote Sensing, Hunan University of Science and Technology under Grant No.E22201; the Agricultural Science and Technology Innovation Program under Grant ASTIP No. CAAS-ZDRW202201; a grant from State Key Laboratory of Resources and Environmental Information System; and the Innovation Capability Improvement Project of Scientific and Technological Small and Medium-sized Enterprises in Shandong Province of China under Grant No.2021TSGC1056.