The Fenxi Coal Mine Area (36°41′20″–37°17′12″N, 111°22′08″–112°21′26″E) belongs to the Huoxi Coalfield andcovers approximately 2,800 km² of land, located in the northern area of the Huoxi Coalfield in Shanxi province, China (Figure 1). The Huoxi Coalfield is one of the six large coalfields (Datong, Ningwu, Hedong, Xishan, Huoxi, Qinshui) of Shanxi province (Su et al., 2017a), which includes the Fenxi Coal Mine Area and Huozhou Coal Mine Area. The landform in the Fenxi Coal Mine Area is complex. The western part of Fenxi is a low mountain area on the east wing of Luliang Mountain. The western demarcate boundaries reach Shuangchi and the Nanyang borderline. The eastern part of Fenxi is the foothills of Huoshan Mountain, bounded by the Huoshan fault. The north part is the Jinzhong rift basin, the north boundaries reach the south of Fenyang City, and the southern border is the Shilin fault and the Fenhe River Valley with flat terrain passes through the middle of the research area. The Fenxi Coal Mine Area belongs to a warm temperate semi-arid continental monsoon climate, with an annual average temperature of 8.6°C and about 180 days in the frost-free period (Su et al., 2017a). Because it is located in the central inland area, the continental climate is relatively obvious, and the temperature difference between day and night is larger (Liu, 2015). The average annual precipitation is 634 mm and rainfall is often in July and August. Slope failures are triggered chiefly by intense and prolonged rainfall in this period (Su et al., 2017a).

The Fenxi Coal Mine Area is on the east margin of the Qi-Lv-He epsilon-type structure arc-fold and different forms of folds compose coal-bearing strata which belong to the Taiyuan Formation (Upper-Carboniferous System) and the Shanxi formation (Lower-Permian System). The average thicknesses of the Taiyuan and Shanxi Formations are about 90 m and 50 m, respectively, which are 810 m and 760 m below ground, respectively. There are abundant coal resources in the study area. The Fenxi Coal Mine Area is in underground mining regions, with blasting and so on. Due to over-exploitation, underground goaf is formed in the mining area, which can easily cause surface damage and form geological hazards. Mine slag includes coal gangue, waste slag, fly ash, and tailings, which damages the geological environment of the mining area and often induces mine geological hazards.

In the research, the main data materials collected are 1) Digital elevation model (ASTER-GDEM), 2) Remote Sensing (RS) images, 3) Geology hazards inventory, 4) Peak earthquake acceleration, 5) Average annual rainfall, 6) Geology map, 7) Road map, 8) Mine area map. Detailed information about the data materials is listed below in Table 1.

Landslides are the movement of a mass of rock, debris, or Earth (soil) down a slope (Wu et al., 2012). The landslide in its widest sense includes colluvial landslides, rock and colluvial collapse, and debris flow, causing gravity erosion of the slope and valley (Qiu, 2012; Wang, 2012; Wu et al., 2012). The formation conditions, inducing factors and movement mechanisms of generalized landslides are diverse, changeable, and complex (Duan, 2010). These kinds of geological hazards are commonly linked in time and space in an interdependent whole, and have the characteristics of chain generation (Ma, 2011). Therefore, the evaluation of these hazards can be represented by a unified whole, and the generalized landslide concept is adopted in the landslide sensitivity evaluation of this paper.

In the Fenxi Coal Mine Area there are 639 geological hazards which include ground collapse (371), unstable slopes (132), colluvial landslides (69), rock and colluvial collapses (36), debris flows (28), ground fissures 2) and land subsidence 1) respectively, obtained from the Geological Environment Monitoring Center of Shanxi Provincial Natural Resources Department. All seven kinds of geological hazards data are converted into certain data formats and finally generate spatial point site data, and shown In Figure 1. The X, Y coordinates of the central point of the geological hazards represent their sites.

According to the definition of landslides mentioned previously, landslides in this paper are including colluvial landslides, rock and colluvial collapses, and debris flows. Unstable slope is a kind of geological hazard site that is prone to landslides. Thus the spatial distribution of landslides in the Fenxi Coal Mine Area with a total of 265 landslide sites has been mapped by applying the remote sensing (RS), geographic information system (GIS), and spatial data analysis method.

Geological hazards destroy the ecological environment, and their occurrence is affected by various internal and external dynamic forces. According to the geological, topographical and survey data of the Fenxi Coal Mine Area, as well as the existing expert experience (Chau et al., 2004; Peart et al., 2005; Domínguez-Cuesta et al., 2007; Xiang et al., 2010), the condition influence factors of landslide hazards including topography, geology, hydrology, land cover, and human engineering activities (Youssef and Pourghasemi, 2021) were selected to explore the mechanism and characteristics in the Fenxi Coal Mine Area. Topography factors are expressed by terrain factors: elevation, slope, aspect, and curvature (plane and profile curvature). ASTER-GDEM with 30 × 30 m spatial resolution was used to extract terrain factors. Geological structural factors include stratigraphic lithology, distance to fault, and peak earthquake acceleration. Many scholars have analyzed the influence of geological structure on the sensitivity of geological hazards and explored it. They believe that strata and strata lithology play a more significant role in ground hazards, and the distance to the fault also directly or indirectly affects them. Stratum lithology is the basis of geological hazards, and some geological hazards occur directly on the stratum. The Hydrology factors contain average annual rainfall and distance to river factors, while land cover factors involve land use type and NDVI. The three human engineering activities factors include distance to road, distance to mined area, and ground collapse density. In the Fenxi Coal Mine Area, the mining area accounts for 55.57% of the total area of the district, and the average density of ground collapse geological hazards points in the Fenxi Coal Mine Area is 0.07 (place/km²), so it is very necessary to construct distance to mined area and ground collapse density factors for the LSP. Through the condition factors and the comprehensive analysis of ArcGIS software, the relationship between landslide hazards and the geological ecological environment was studied, and its sensitivity was classified and evaluated. In the Fenxi Coal Mine Area the whole landslides condition factors involved are described and mapped below:

RF is the most commonly used machine learning algorithm. It is a combined model that integrates multiple decision trees. By combining multiple weak classifiers, the final result is voted or averaged, which makes the result an overall model with high accuracy and generalization performance (Maxwell et al., 2020). The working principle of RF is shown in Figure 3. RF involves two concepts: 1) random sampling of data points; 2) Segmentation of nodes based on feature subsets. Random sampling: Each tree is trained on a sample of data points drawn at random. These samples are drawn repeatedly; random subset of features: At each node, the decision tree will consider segmentation based on a part of the feature, and the number of this part of the feature is the square root of all of the elements features. The best conditions for predictors are given by log₂ (M+1), where M represents the number of inputs to the algorithm, and the mean square error is given by (Sevgen et al., 2019): ε = ( v 1 − v 2 ) 2 (1)

Among them, ε represents the mean square error, V₁ is the variable of the observation data, and V₂ is the result variable (Sevgen et al., 2019). The calculation formula of the mean is: S = 1 K ∑ K t h v 2 (2)

In the formula, S is any forest prediction value, while K is used to a single tree in RF, V₂ is the result variable too. Through this algorithm, on the basis of determining the simple tree set and random predictor variables (Band et al., 2020), edge functions are defined.

C5.0 is a decision tree generation algorithm (Tobi and Duncan, 2019). It divides the sample data in the light of the field that provides the maximum attribute of information gain rate as a critical method to determine the arithmetic performance, at the same time cuts and merges the leaf nodes of the decision tree to raise the classification accuracy, ultimately determine the optimal threshold for every leaf. The core of the C5.0 algorithm is to select the characteristic variables of each branch by using the speed of information entropy reduction (Guo et al., 2021). The calculation formula of information entropy is as follows (Liu et al., 2017): H ( U ) = ∑ i P ( u i ) log 2 1 P ( u i ) (3)

In the formula, u_i (i = 1,2, …r) represents information, and P (u_i) represents the probability of occurrence of information ui (i = 1,2, …r). When encountering too much or insufficient sample data, the C5.0 will automatically eliminate and adjust the weights to obtain an optimal decision tree model. C5.0 possess the advantages of high reliability, fast running speed, small memory usage, and high fault tolerance (Tobi and Ducan, 2019; Guo et al., 2021).

On the basis of the structural risk minimization principle and VC dimension theory of statistical learning theory, Vapnik et al. developed another machine learning method: SVM. The SVM model is based on a variety of basic functions to transform linearly inseparable data into high-dimensional space, and find hyperplanes in that high-dimensional space to realize linearly separable data patterns (Chen et al., 2016; Su et al., 2017b; Liu et al., 2017). SVM is able to turn a nonlinear problem into a linear one in a high-dimensional space by nonlinear transformation, and then find the optimal classification surface in the transformed high-dimensional space (Su et al., 2017a). SVM skillfully solves the problem of mapping from low dimensional input space to high dimensional feature space by introducing kernel function (Huang et al., 2018; Guo et al., 2021). In this paper, the kernel function of the Radial basis function (RBF) is used, and the algorithm formula is as below: K ( x i , x j ) = e − γ ( x i − x j ) 2 (4)

Among them, the γ of the kernel function is a parameter, which needs to be optimized when building the model to increase the fitting accuracy.

After determining the landslide inventory map and influencing factor maps (Guo et al., 2021), RF, SVM, and C5.0 algorithms were integrated to generate the final landslide susceptibility map. The Fenxi Coal Mine Area has a total of 265 landslides, which equaled to non-landslide sites that were built, and changed to pixels according to (Su et al., 2017a), for the sake of offering the essential knowledge about stable or unfavorable conditions of landslide occurrence (Guo et al., 2021). Through the Geostatistics analysis means the pixels are separated entirely into two parts in a random manner: 1) The first part contains 80% of the data, participated in the model training stage; 2) The second part is the validation data set, including the remaining 20% of the data adopted to verify these three models and confirm their accuracy.

The three measures of Accuracy, Specificity, and Sensitivity were adopted to evaluate the performance after applying a dataset of 5-fold cross-validation. Also, the Receiver Operating Characteristics (ROC) of each model were plotted, and the Areas Under the ROC Curves (AUC) of every model were obtained. An SLP map has the ability to predict future landslide sites. RF, SVM, and C5.0 are used to model and verify in the two (training and verification) phases, which can be adopted to obtain the landslide sensitivity of all pixels in the Fenxi Coal Mine Area. Jenks Natural Breaks algorithm was used to divide the probability range of landslide into five grades: very low, low, medium, high, and very high, and the three landslide susceptibility maps were obtained. Two aspects are listed below in which the landslide sensitivity zoning performance is able to be illustrated: 1) the surveyed landslide sites mainly scatter in the very high and high LSP zone, and it shows that the performance accuracy of the LSP zone is relatively higher; 2) the sites in the very high and high sensitivity zone in all the surveys ought to account for lower portion correspondingly, which can effectively reduce the redundancy and improve the LSP accuracy (Gokceoglu et al., 2005; Su et al., 2017a).

In this paper, a total of 15 condition factors–slope, elevation, aspect, curvature (plane curvature, profile curvature), lithology, distance to fault, peak earthquake acceleration, land use type, NDVI, distance to river, annual average rainfall, distance to road, distance to mined area and ground collapse density—were selected as the susceptibility evaluation indices. In order to ensure the objectivity and independence of the evaluation indices, we adopted the Pearson correlation analysis method to analyze the correlation among the 15 indicators. The SPSS software was used to import the matrix of 15 indices factors, and the correlation analysis was carried out. The correlation between condition factors is displayed in Table 2, and Table 2 indicates that all indices condition factors meet the requirements of independence. Combined with the actual situation of the Fenxi Coal Mine Area, we use these 15 condition factors as the evaluation indices of LSP in the study area.

(In Table 2 the 15 Landslide condition factors (elevation, slope, aspect, plan curvature, profile curvature, lithology, distance to fault, peak earthquake acceleration, distance to river, annual average rainfall, land use type, NDVI, distance to road, distance to mined area and ground collapse density) are shorted and represented by Elev, Slop, Aspe, Pl-c, Pr-c, Lith, Faul, Peak, Rive, Rain, Land, Ndvi, Road, Mine and Coll, respectively).

For the use of the three models, the input and output variables, the training and testing data are all determined as shown in Section 2.4. Cross-validation estimation of the predictive performance of a model is a crucial step in predictive modeling, and spatial cross-validation is recommended for spatial data, which may be subject to spatial autocorrelation (Su et al., 2017a), so the 5-fold cross-validation mean is adopted to calculate these LSP indices. In the training process of C5.0, the redundant nodes of the tree are pruned along with the tree growth, and the child nodes are created 10 times (Huang et al., 2020c), and the other model parameters are given as default. During the training process of RF, the maximum number of nodes is selected as 10,000, the maximum tree depth is set to 10, and the Minimum node size is 5. The model will stop building when the accuracy is no longer improved. All the parameters of SVM, Epsilon is 0.1, cost constant is 10, and RBF-γ, γ, and degree is 0.1, 3, respectively. Through cross-validation of the model, the confusion matrix was obtained, and based on this, the fitting accuracy is calculated and shown in Table 3.

Table 3 indicates that the average accuracy rates of C5.0, RF, and SVM in the training stage are 97.36, 93.87, and 84.24%, respectively. The standard deviation of C5.0, RF, and SVM in the training stage is 2.85, 3.11, and 0.81. C5.0 has the highest accuracy and lowest standard deviation in the training stage, so C5.0 performs better in the training stage. In the validating phase, their average accuracy rates were 69.81, 65.47, and 67.17%, respectively. Therefore, C5.0 has the highest accuracy rate in the verification phase, followed by the SVM accuracy rate, and the RF accuracy rate is the lowest. Their standard deviations in the validating phase are 3.34, 2.17, and 3.97, respectively. Obviously, RF has the lowest standard deviation in the verification phase, the standard deviation of C5.0 is the second, and the standard deviation of SVM is the highest.

Using a 5-fold cross-validation dataset for C5.0, RF, and SVM algorithms the ROC curves and their corresponding AUC values of the LSP are illustrated in Figure 4. The results of AUC in the training stage are shown in Figure 4A. The estimation of the training set is always too optimistic (Brenning, 2005; Su et al., 2017a). In accordance with AUC obtained from the training stage, wholly, the AUC indices values are indeed much higher than 0.9, indicating that C5.0, RF, and SVM algorithms completely obtain a successful performance. C5.0 has much better prediction results (0.99), compared with RF (0.96) and SVM (0.92). Prominently, C5.0 is much more successful than RF and SVM in the training stage. This means that the current three models (C5.0, RF, and SVM) are all capable of solving complex questions.

In the validating stage, the values of AUC correspondingly are illustrated in Figure 4B. The values of AUC range from 0.71 to 0.80 in the validating stage, the same as in the training stage, C5.0 is the highest one with 0.80, followed by RF, and SVM which both have the performance of 0.71. With an AUC value of more than 0.7, all the three models are capable of modeling the LSP in Fenxi Coal Mine Area in this study. Nonetheless, C5.0 appears to be more accurate and promising than RF and SVM.

The same as the fact that different models have different performances, different condition factors cannot make equal contributions to the evolution of regional landslides (Chen et al., 2018; Guo et al., 2021). The contributions of condition factors in the Fenxi Coal Mine Area are determined through calculating the mean value of relative contributions of each condition factor under all the three different models (Guo et al., 2021), and the contributions of each cross variable condition factors of these fitting algorithms are shown in Figure 5. Figure 5 mainly shows that the median contributions of the condition factors of the variables are classified as distance to road, lithology, profile curvature, ground collapse density, land use type, average annual rainfall, slope, distance to mined area, NDVI, aspect, peak earthquake acceleration, distance to river, and so on in descending contribution order, from highest to lowest. Hence, it is obvious that distance to road, stratum lithology, profile curvature, and ground collapse density are principal condition factors in the Fenxi Coal Mine Area for the evolutions of landslides. The distance to mined area factor has a medium contribution, and the ground collapse density factor plays an obvious role in the occurrence of geological hazards.

The study area contains 3,597,447 pixels, converted into point type and mapped by ArcGIS platform. Using Jenks Natural Breaks algorithm, the three landslide susceptibility indices were reclassified into five susceptibility levels, shown in Table 4.

In accordance with the Classification standard for susceptibility of landslides in the Fenxi Coal Mine Area, Three landslide sensitivity maps are obtained and converted into a grid format, and the three LSP maps using the three models are plotted in Figure 6. Figure 6 indicates that the three algorithms consistently have given the northern zone in the Fenxi Coal Mine Area a low or moderate record score of LSP. It is realistic that this zone has a lower elevation, and rarely slope instabilities generate in Jinzhong Basin. On the contrary, areas with the value of ground collapse density increasing was given higher scores in LSP in the Fenxi Coal Mine Area.

In order to compare the three spatial LSP indices, the characteristics of landslides distribution and their area percentage under each susceptibility range are shown in Table 5.

It is obvious from Table 5 that regions with high and very high landslide sensitivity degrees possess 193, 173, and 176 landslide sites, when considering the area percentage, they also possess 17.55, 23.45, and 23.44% for the three models for C5.0, RF and SVM respectively in the Fenxi Coal Mine Area. Because the surveyed landslide hazard sites are mostly distributed in the regions with high landslide sensitivity, and in all the surveys, the sites with high sensitivity grades account for a very low portion (Gokceoglu et al., 2005; Su et al., 2017a). Therefore, C5.0 is the best one of all the models for the LSP algorithms.

The three typical models are widely used in many studies and can well embody the basic features of their corresponding model types (Huang et al., 2020c; Guo et al., 2021). RF has the advantages of fewer restrictions on variables involved in the evaluation, no need to consider the data scale and data distribution, high computational efficiency, high precision, and low debugging cost compared with other deep learning models (Su et al., 2017b; Huang et al., 2018). SVM is capable of solving some matters with nonlinearity, small samples, over-learning, dimensionality curse, and local minima, and has strong generalization ability. But for many types of problems, SVM is not efficient (Wu et al., 2014). C5.0 possesses some superiorities in modeling the SLP (Alkhasawneh et al., 2014; Park and Lee, 2014; Wu et al., 2014). C5.0 is a good model which is easy to understand and explain, training needs less data, and the modeling processes includes tree nodes grown, tree nodes pruning, feature selection, and so on. The other machine learning models usually need the advantages of data normalization, but the results of C5 are easy to overfit. Hence, through the comparison of these algorithms, C5.0 can be understood to have the ability to overcome the shortcomings that the traditional machine learning models possessed. The C5.0 is the best performing model of all machine learning algorithms.

The accuracy of the LSP model is decided by the algorithm selected (Su et al., 2017a; Huang et al., 2020c). The average precision of C5.0, RF, and SVM in the training stage is 97.36, 93.87, and 84.24% respectively, and their accuracy in the validating stage is 69.81, 65.47, and 67.17%. The standard deviation represents the stability of the model, and the higher the accuracy is, the more unstable the algorithm is. The standard deviation of C5.0, RF, and SVM is 2.85, 3.11, and 0.81 in the training stage, and is 3.34, 2.17, and 3.97 in the validating stage respectively. Therefore, C5.0 is more moderate for LSP in this study. The algorithms which have a high AUC value usually possess higher and more successful prediction indices (Huang et al., 2020c). The ROC curves of the three models suggest that in the training stage C5.0 has a much higher prediction index (0.99) than that of RF (0.96) and SVM (0.92), while in the validating stage C5.0 has the highest AUC with 0.80, followed by RF (0.71) and SVM (0.71), the same as the performance in the modeling process. Therefore, considering the fitting accuracy, ROC curve, and AUC value of C5.0, RF, and SVM, C5.0 is more moderate than the other two algorithms.

The attribute data in Fenxi Coal Mine Area is introduced into the C5.0, RF, and SVM to obtain the landslide susceptibility indices. Taking ArcGIS as the sensitivity indices, Jenks Natural Breaks algorithm is adopted to divide the probability indices of landslide into five levels that are very low, low, medium, high, and very high in indices ascending order, and three landslide sensitivity maps are obtained respectively. The results indicated C5.0, RF, and SVM have 193, 173 and 176 landslide sites in the high and very high areas respectively, and the high and very highly prone areas account for 17.55, 23.45, and 23.44% of the study area. Because the surveyed landslide hazard sites are mostly scattered in which the high and very high landslide sensitivity levels are distributed, while in the whole process, the landslide sites with high sensitivity levels account for a low proportion (Su et al., 2017a). Therefore, C5.0 is the best of all the models. These three models also predict that the landslide sensitivity of the Jinzhong basin is low or moderate, and predict that the landslide sensitivity of the mountains lies in the central section, the northeast part, and the east part of the Fenxi Coal Mine Area and increases with the value of ground collapse density.

Generally, the factors such as Topography, Hydrology, Geology, land use type, and so on are extensively accepted as condition factors in many LSP models (Su et al., 2017a). Distance to road and lithology factors are the two important factors. The highest contribution of distance to the road may be that the road constructions have changed the slope, formed an escarpment, and led to slope instability. On the mechanical properties, the lithology factor can affect the slope instability, it has been proved that the hard and dense rock mass hardly has any sensitivity to slope instability (Huang et al., 2020a).

Profile curvature and ground collapse density are relatively important factors too. The curvature factors represent the unevenness of the land surface. The greater the unevenness, the more uneven the force of the slope, and the more likely geological hazards are to occur (Su et al., 2017a). The ground collapse density theoretically plays an obvious role in the occurrence of geological hazards. Ground collapse significantly affects the stability of the surface of the mining area. Abundant ground collapse destroys the surface morphology and stability, which is closely related to geological hazards and ecological environment sensitivity in mining areas.

Then the factor contribution importance of the land use type, average annual rainfall, slope, distance to mined area, NDVI, Aspect, peak earthquake acceleration, and so on are listed in descending order of contribution. It is obvious that the distance to mined area factor has a medium contribution to all the models.

In the Fenxi Coal Mine Area the mining area accounts for 55.57% of the total area of the district, and the average density of ground collapse geological hazards points is 0.07 (place/km²). Mining disturbance has great damage to the surface, and ground collapse density affected the occurrence of geological hazards. Thus the mining disturbance such as the underground mined area and ground collapse have prominently affected the slope instability of the Fenxi Coal Mine Area.