Investigating machine learning and statistical approaches for landslide susceptibility mapping in Minfeng County, Xinjiang

Yangyang Jiao; Daozhu Xu; Qiang Wang; Lei Wang

doi:10.1515/geo-2025-0906

40% Rabatt

auf Fachbücher bei De Gruyter Brill *

Artikel Open Access

Investigating machine learning and statistical approaches for landslide susceptibility mapping in Minfeng County, Xinjiang

Yangyang Jiao , Daozhu Xu , Qiang Wang und Lei Wang

Veröffentlicht/Copyright: 5. November 2025

Veröffentlicht von

Veröffentlichen auch Sie bei De Gruyter Brill

Manuskript einreichen Informationen für Autor*innen

Aus der Zeitschrift Open Geosciences Band 17 Heft 1

Abstract

Landslide susceptibility assessment in arid mountainous regions requires specialized modeling approaches. This study, combining the information value (IV) modeling and machine learning, develops a coupled model approach for Minfeng County, Xinjiang, that a complex arid zone with frequent landslides. From the ten influencing factors, seven key factors were identified through factor covariance and correlation studies, so as to construct the landslide susceptibility evaluation index system. On this basis, using 135 landslide samples and combining the output of the information value (IV) model with four machine learning algorithms–support vector machine (SVM), logistic regression (LR), random forest (RF), and artificial neural network (ANN)–we constructed four coupled models (IV-LR, IV-ANN, IV-SVM, and IV-RF) for landslide susceptibility evaluation. Critical results are as follows: (1) proximity to rivers/roads and vegetation density (NDVI) dominate landslide triggers and (2) all models showed high accuracy (area under curve [AUC] > 85%) with 7:3 training:testing validation and the IV-RF model achieved optimal high-susceptibility zone delineation (accuracy = 82.71%; AUC = 0.8945). This method provides a technical reference for landslide disaster prediction, prevention, and mitigation in arid mountainous areas of Xinjiang.

Keywords: landslide risk assessment; hybrid modeling, information value model; random forest; logistic regression; artificial neural network; support vector machine; arid mountainous regions; Xinjiang

1 Introduction

Landslides are a type of geological disaster. They are widely distributed, highly sudden and destructive, seriously affecting the social and economic development of regions [1]. Regional geological environment is the fundamental condition for the formation of landslides. Human engineering activities, rainfall, etc. [2], are the inducing factors of landslides. Precise evaluation of landslide susceptibility has always been a research hotspot in landslide risk assessment and management [3]. Due to the complexity of the terrain and landforms and the frequent occurrence of extreme weather in the central mountainous area of Minfeng County, Xinjiang Uygur Autonomous Region, landslide disasters occur frequently, which have the distribution characteristics of “many points and wide area.” Therefore, it is necessary to conduct an assessment of the susceptibility of landslide disasters in Minfeng County, providing a scientific basis for government decision-making, early prediction, and disaster prevention and mitigation.

Statistical analysis methods cannot explain the nonlinear relationships among various disaster-causing factors [4]. Machine learning models have strong learning capabilities but are prone to overfitting. Some scholars have attempted to use statistical analysis coupled with machine learning models to evaluate the susceptibility of landslides, and the results have proved that it can effectively improve the accuracy of landslide susceptibility evaluation. Saha et al. [5]. compared the results of the AHP-SVM single hybrid model under different function algorithms when studying landslide disasters in the Himalayan region. Youssef et al. [6] integrated logistic regression (LR) and artificial neural network (ANN) in their research on the southern Sinai Peninsula, and the integrated model s prediction results were superior to those of the individual models; the research conducted by Gu et al. [7] in the river valley area concluded that the sample quality of non-landslide sampling differed significantly when using different machine learning methods. Saha et al. [8] conducted a landslide susceptibility evaluation in the study area in the Himalayan region by combining GIS, statistical methods, and simple machine learning methods, and the accuracy was satisfactory. Abdelkader and Csámer [9] conducted an evaluation study on landslide susceptibility in the arid regions of Egypt by using machine learning methods. The results demonstrated that machine learning can enhance the accuracy of susceptibility evaluation in arid areas [10,11].

This study aims to construct a transferable susceptibility assessment system for arid mountainous areas, quantitatively compare the performance of IV-LR, IV-ANN, IV-SVM, and IV-RF models, and determine the optimal zoning method for landslide risk management in Minfeng County, with the potential to be applied to similar arid areas.

2 Materials and methods

2.1 Study area

Minfeng County is located in the northern foothills of the Kunlun Mountains, the southern edge of the Taklamakan Desert in Xinjiang. The county is situated in longitude 82°22′85°55′E and latitude 35°20′–39°29′N, with a total area of 5.67 × 10⁴ km² [12,13]. The study area belongs to the typical temperate desert row climate. The study area has a typical temperate desert climate, with a mean annual precipitation of 30.5 mm and that heavy summer rains are prone to landslides [14]. The topography of the area is high in the south and low in the north. The landforms are made up of three major topographic units, namely, Kunlun Mountain in the south, the alluvial plains in the north, and the deserts, and the geologic structure is complex and varied, which mainly includes magmatic rocks, sedimentary rocks, and metamorphic rocks [15]. The geological structure is complex and diversified, mainly including magmatic rocks, sedimentary rocks, and metamorphic rocks. The complex geological structure, the development of ruptures and folds, and the relatively active neotectonic movement have exacerbated the fragility of the geological environment [16]. Frequent landslide disasters occur in the southern mountainous areas, particularly on both sides of river valleys, steep slopes, and areas around frequent human engineering activities like road construction and mining. According to the statistics, during the period of 2020–2023, a total of 2,765 times of landslide disasters occurred in 135 landslide investigation sites in the region, as shown in Figure 1.

Figure 1

Location map of the study area.

2.2 Data sources

According to the requirements of field investigation and evaluation accuracy [17], the data of each index factor were collected, and the data source is shown in Table 1.

Table 1

Data information

Data type	Indicator factors	Data sources	Data description
Digital elevation model	Slope, aspect, elevation	ASTER-GDEM (30 m) https://www.gscloud.cn/	Extract elevation, slope, aspect
Remote sensing image	NDVI	Landsat 8 (30 m) http://www.resdc.cn	Extraction of normalized vegetation index
Geological map	Geotechnical type, distance from fault, peak ground acceleration (PGA), landslide points	Geological survey department (1:500,000)	Extraction of stratigraphic lithology, distance to fault, and PGA
Road, river system data	Distance from rivers, distance from roads	Natural resources department (1:500,000)	Extraction distance from roads and rivers
Precipitation data	Precipitation	National Weather Data Center (1 km) http://www.ngac.org.cn/	Extraction of precipitation data

2.3 Landslide inventory

Landslide inventory mapping is an important part of landslide susceptibility mapping [18,19]. To model landslide susceptibility zoning, a landslide inventory is first required, and a landslide inventory map is generated by making a landslide inventory of the study area, as shown in Figure 2. A total of 135 landslides occurred in this study area, the non-landslide points were randomly generated in the non-landslide area according to 1:1, and 70% were randomly divided in a ratio of 7:3 as a training set and 30% as a test set for model validation [20].

Figure 2

Landslide inventory map of the study area.

2.4 Landslide influencing factors

The occurrence and development of landslides are jointly influenced by factors such as topography and geomorphology, geological environment, hydrological conditions, and human activities [21,22,23]. In the process of regional landslide susceptibility evaluation, it is crucial to select environmental evaluation factors reasonably. In this study, sufficient field investigation and data collection were conducted in the Minfeng County region of Xinjiang, and combined with the experience of previous studies, ten indicators of slope gradient, slope direction, elevation, geotechnical body type, distance from faults, distance from rivers, NDVI, distance from roads, average annual precipitation, and peak ground acceleration (PGA) from earthquakes were selected, as shown in Figure 3.

Figure 3

Landslide influencing factors: (a) elevation; (b) slope; (c) aspect; (d) geotechnical type; (e) distance from fault; (f) distance from main river; (g) distance from roads; (h) NDVI; (i) precipitation; and (j) PGA.

2.5 Methodology

In this study, the GIS-based information value (IV) statistical analysis model coupled with support vector machine (SVM), LR, ANN, and random forest (RF) methods based on Python programming language for predicting landslide susceptibility maps [24], in addition to the validation and comparison using receiver operating characteristic (ROC) curves and model evaluation metrics, were used, as shown in Figure 4.

Figure 4

Overall methodological representation of the work.

2.5.1 IV

The IV method is a statistical prediction method, which has a basis for partitioning the landslide susceptibility in the region by counting the amount of information provided by each factor in the historical landslide disaster [25]. The larger the value of the information quantity, the greater the influence of the indicator factor on landslides, and the greater the likelihood of landslide disasters when zoning is carried out. In this article, the probability of geohazard occurrence is measured comprehensively through the statistical specific evaluation unit information volume, the information volume of different influence factors on the occurrence of geohazards is calculated, the contribution weights of the information volume of different factors are derived, the total information volume is obtained through weighted summation, this is used as the input parameter of machine learning, and the method of calculating the IV value is shown in equation (1) [19]:

(1) IV ( i , j ) = In ( N ( i , j ) / N ) ( S ( i , j ) / S ) ,

where IV_(i,j) denotes the informativeness value of the ith factor at level j, informativeness value provided by indicator factor (or range) i for the occurrence of landslide hazards in the region; N _(i,j) denotes the number of hazard sites of the ith factor at level j for the specific grading interval of the indicator factor (or range); N is the total number of hazard sites in the study area; S _(i,j) denotes the area occupied by the ith factor at level j for the specific grading of the indicator factor (or range) i, km²; and S is the total area of the study area, km².

2.5.2 SVM

The SVM model is a new generation of algorithms that originated from risk minimization. The basic principle of SVM is to map the input vector data (nonlinearly divisible) into a higher-dimensional feature space by means of a kernel function (nonlinear mapping) [26] and then to find the classification-optimal hyperplane that can efficiently separate the two classes of data in this space [11]. SVM is a kind of supervised classification for binary classification integrating the kernel function, the technique of maximal interval hyperplane, and the technique of maximum interval hyperplane. It is suitable for classifying data with small samples, nonlinearity, and high dimensionality. When SVM deals with a two-class classification problem, the training sample set X _i (i = 1, 2, …, n) is divided into two classes, denoted as Y _i = ±1. SVM has to find an n-dimensional hyperplane (plane or surface) that distinguishes the two classes of data at a maximum interval and requires that the distance from the separated data points to the hyperplane is maximum. The mathematical expression and constraints for the hyperplane H are given in equations (2) and (3) [27], respectively.

(2) min 1 2 ∥ H ∥ 2 ,

(3) Y i ( ( H ⋅ X i ) + a ) ≥ 1 ,

where ∥ H ∥ 2 denotes the norm of the normal vector of the plane and a represent a scalar and a scalar product, respectively. Introduce Lagrange multipliers λ _i and construct the auxiliary function as equation (4) [28]:

(4) L = 1 2 ∥ H ∥ 2 − ∑ i = 1 n λ i ( Y i ( ( H ⋅ X i ) + a ) − 1 ) .

Assume the partial derivatives of L with respect to H and a are both zero, deriving equations (5) and (6) [29,30]:

(5) H = ∑ i = 1 n λ i Y i X i ,

(6) ∑ i = 1 n λ i Y i = 0 .

Substitute equations (5) and (6) into equation (4), obtaining equation (7) [31]:

(7) L = ∑ i = 1 n λ i − 1 2 ∑ i = 1 n ∑ j = 1 n λ i λ j Y i Y j ( X i ⋅ X j ) .

2.5.3 LR

The LR model is a linear regression method, the model is based on binomial classification [32], which can effectively analyze the accuracy of qualitative variables, and the data of a single environmental impact factor are used as the independent variable and the dependent variable is whether the landslide disaster occurs when the LR model is used to evaluate the susceptibility of landslides. Among them, the landslide occurrence value is 1, and the landslide non-occurrence value is 0. In general, the LR model is expressed as in equation (6) [33]:

(8) L ( P ) = ln p 1 − p = C 0 + C 1 x 1 + C 2 x 2 + ⋯ + C n x n ,

where p and 1 − p are the probability of landslide occurrence and non-occurrence respectively, L ( P ) is the objective function of the probability of landslide occurrence, and C 0 is a constant, which refers to the logarithm of the ratio of p to 1 − p when there is no influence of any factors; [x ₁, x ₂, …, x _n] is the set of influencing factors of independent variables, and [C ₁, C ₂, …, C _n] is the LR coefficient, that is, the estimation parameters of [x ₁, x ₂, …, x _n].

2.5.4 ANN

ANNs are shallow machine learning algorithms that have been widely used in various fields of research [34]. There is a complex nonlinear relationship between the factors of landslide susceptibility assessment, and ANN is one of the effective tools used to solve this nonlinear problem. The ANN system consists of three parts: the input layer, the hidden layer, and the output layer [35]. In the process of network learning training, first, the connection weights between the factors are assigned according to the established input values and output values and, second, the model selects the appropriate activation function for repeated training, so as to readjust the connection weights, so as to correctly evaluate the nonlinear relationship between the factors, so that the output mean square error value is smaller, and the accuracy of the model is improved accordingly. The ANN model is expressed as in equation (7) [36]:

(9) y i = f ∑ w i j x i + b j ,

where f is the activation function; w i j is the weight between neurons i and j; and b j is the bias of neuronal j.

2.5.5 RF

RF model is a model based on multiple decision trees; the basic principle is to create multiple decision trees for different factor datasets, train and test the samples using multiple decision trees, and then vote on the results obtained from multiple decision trees to get the final results of the RF model [37]. The RF model is randomly selected in the training set as well as the split attributes are randomly selected, which prevents the model from overfitting and improves stability.

3 Results and analysis

For the factors selected for the study, the correlation and covariance were used to obtain the final critical evaluation factors, which were combined with the IV model to complete the calculation of the information content and the ranking of the weights, and the landslide susceptibility evaluation maps were obtained by combining the machine learning model with the landslide and non-landslide (1:1) data for training as well as the comparison of the ROC curves of the four methods (Figure 4).

3.1 Indicator factor multicollinearity and correlation analysis

3.1.1 Multicollinearity analysis

The independence between the influencing factors needs to be detected before modeling landslide susceptibility zoning to exclude the influencing factors that cause multicollinearity [38]. In this article, the variance inflation factor (VIF) and tolerance (TOL) are used to detect the multicollinearity between the factors, and if the value of VIF is larger and the value of TOL is smaller, it means that the multicollinearity between the factors is stronger. Generally speaking, if VIF > 4 and TOL < 0.25 indicate that the corresponding influence factor will cause serious multicollinearity problems, then the factor is not involved in the subsequent modeling (Table 2).

Table 2

Multicollinearity detection table

Indicator factors	TOL	VIF
NDVI	0.952	1.051
Elevation	0.137	7.288
Geotechnical type	0.685	1.46
Slope	0.716	1.396
Aspect	0.941	1.063
Distance from fault	0.627	2.058
Distance from river	0.952	1.05
Distance from road	0.580	2.085
PGA	0.246	4.241
precipitation	0.145	6.893

In this article, linear regression analysis is done for the influence factors of the input data, respectively, and two sets of VIF and TOL values are obtained. As can be seen from Table 2, the top three VIF and TOL values of the two indicators are elevation, precipitation, and PGA, respectively. The VIF and TOL values of these three factors are within the range of VIF > 4 and TOL < 0.25, so there is a problem of multiple covariance between these three influencing factors, and therefore, enter the exclusion zone.

3.1.2 Correlation analysis

The correlation test of 10 indicators using the Pearson correlation [39] coefficient is shown in Figure 5. Among them, there are three groups of indicator factors with strong correlation, which are highly significant, namely elevation and precipitation (0.889**), elevation and PGA (0.311**), and PGA and precipitation (0.241**). Elevation, precipitation, and PGA were excluded in conjunction with the results of the previous covariates.

Figure 5

Evaluation indicator correlation matrix. Note: *** indicates a significant correlation at the <0.05 level (two-tailed) for P-values, where P-values indicate the level of significance of the Pearson correlation coefficient.

3.2 Landslide susceptibility mapping and analysis

3.2.1 IV

Before calculating the IV value, each environmental impact factor was reclassified using the natural discontinuity grading method, and the results of the grading criteria, IV value, and IV value ranking calculation for each impact factor are shown in Table 3.

Table 3

IV of each impact indicator

Factor	Rank	IV	Sort	Factor	Rank	IV	Sort
Slope	0–5	−0.7604	32	Aspect	Plane	—	38
	5–15	0.0749	18		North	−0.1851	25
	15–25	0.5252	9		North east	−0.0709	23
	25–35	0.2069	17		East	−0.0133	19
	>35	−0.3603	30		South east	−0.4476	31
NDVI	<0.2	1.4869	3		South	0.4988	12
	0.2–0.4	0.5021	10		South west	0.3723	13
	0.4–0.6	0.2988	14		West	−0.0320	20
	0.6–0.8	−0.2106	28		North west	−0.2408	29
	>0.8	−0.0414	21	Geotechnical	Mudstone	−1.6926	35
Distance from road	>1,600	−0.1474	24		Sandstone	−1.5317	34
	1,200–1,600	−0.0707	22		Loss	0.2285	15
	800–1,200	0.6144	8		Sandy mudstone	1.5417	2
	400–800	1.4169	5	Distance from river	>1,600	−2.9559	37
	0–400	1.6140	1		1,200–1,600	−0.8884	33
Distance from fault	>1,600	−0.1957	26		800–1,200	−1.7886	36
	1,200–1,600	0.5009	11		400–800	−0.1981	27
	800–1,200	0.6206	7		0–400	1.4816	4
	400–800	0.2165	16
	0–400	0.8216	6

It can be seen that the largest IV value is according to the road (0–400 m), the smallest IV value is according to the river (>1,600 m), and the rest of the top 5 are vegetation index <0.2, sandy mudstone, according to the river (0–400 m) and according to the road (400–800 m) grading factors. The main factors affecting landslide susceptibility in the area are still sparsely vegetated areas, areas with unstable geotechnical mixes, areas near rivers prone to water erosion, and areas with high human activity.

The standardized values can be calculated using the IV values within different levels of each evaluation factor and the number of disaster points, and the calculation results of the standardized weight index values for each evaluation factor can be obtained, as shown in Table 4.

Table 4

Numerical calculation results of standardized weight index of each evaluation factor

Factor	IV_weight	Factor	IV_weight
NDVI	2.0642	Geotechnical	0.0000
Distance from fault	0.0315	Slope	0.1150
Distance from road	0.8016	Distance from river	1.7823
Aspect	0.0000

The largest weight index is the NDVI, the smallest is the aspect, and the top 3 contributors are the NDVI, the distance from rivers, and the distance from roads.

3.2.2 Model training

The susceptibility of landslides in Minfeng County was evaluated. The spatial resolution of the slope prediction unit was 30 m. Outside the 135 landslide points, the following measures were taken: (1:1) construct non-landslide points, with the distance between non-landslide points being greater than 400 m and the distance between any two non-landslide points being greater than 200 m. Simultaneously generate negative landslide sample points, assign values 0 and 1 to the positive and negative sample points as input variables of the model, and reassign the original values of the environmental factors as IV values as output variables of the model. The landslide sample point data were randomly divided into the training set and the test set in a 7:3 ratio. The IV values of the seven influencing factors in the entire study area were input into the model. The prediction results were divided into five intervals of extremely low, low, medium, high and extremely high vulnerability according to the natural discontinuity point method. The above training process was implemented using Python code.

In this study, four machine learning models were selected for the zoning of landslide susceptibility [40]. The accuracy of the models not only depends on the learning algorithm but is also affected by hyperparameters and feature (factor) selection. Therefore, each model needs to be optimized, including hyperparameter optimization and feature selection. Hyperparameter optimization is to select the optimal hyperparameter values based on evaluation indicators [41]. The commonly used hyperparameter optimization methods at present include the random search method, the grid search method, and Bayesian optimization. This study employed a grid search algorithm to determine the optimal hyperparameter values and specifically utilized Python to optimize the machine learning model with the accuracy of cross-validation. Table 5 shows the final hyperparameter values adopted by the five machine learning models.

Table 5

The optimal hyperparameters of four machine learning models

Models	The best parameters of hyperparameter
IV-SVM	{“C”: 10, “gamma”: 0.1, “kernel”: “rbf”}
IV-LR	{“C”: 100, “penalty”: “l1,” “solver”: “liblinear”}
IV-RF	{“max_depth”: None, “min_samples_leaf”: 4, “min_samples_split”: 10, “n_estimators”: 100}
IV-ANN	{“activation”: “tanh,” “hidden_layer_sizes”: (50), “learning_rate_init”: 0.001, “solver”: “adam”}

3.2.3 The results of landslide susceptibility mapping

Four IV model coupled machine learning models were constructed, namely IV-LR, IV-RF, IV-SVM, and IV-ANN models. The prediction results are divided into five intervals of extremely low, low, moderate, high, and extremely high susceptibility according to the natural discontinuity point method.

As can be seen from Figure 6 and Table 6, the prediction results of different coupling models have certain similarities. The landslide susceptibility zoning map has the following characteristics (Figure 7):

The areas prone to high and extremely high landslides are mainly distributed within a distance of about 400 m from water systems. The softening and scouring effect of water systems on riverbank slopes can reduce the strength of rock masses, which is conducive to the development and formation of landslide disasters.
The areas prone to extremely low and low landslides are mainly distributed in the northern and southern regions of the study area. However, the predicted areas vary across different models. Under the prediction of the IV-RF model, the area of the areas prone to extremely low and low landslides is the largest, accounting for 68%, followed by the IV-ANN model, and the IV-SVM model predicts the smallest area.
Comparison of different models shows that landslide disasters almost fall into high and extremely high-risk areas. Among them, the IV-RF model predicts that the proportion of disaster points falling into high and extremely high-risk areas is the highest at 88.89%, which also indicates that this model has higher prediction accuracy and is more suitable for the evaluation of landslide disaster susceptibility in this region.
Among the evaluation results of each model, the landslide disaster-prone areas in Minfeng County are mainly distributed in the middle. The areas of high and extremely high grade prone areas are relatively small compared to those of low and extremely low grade prone areas, which are closer to roads and densely populated with human activities, thus being more conducive to the development of landslide disasters.

Figure 6

Landslide susceptibility map of (a) IV; (b) weight of evidence (WOE); (c) LR; and (d) ANN.

Table 6

Table of area shares of different model susceptibility zones

Model	Spatial distribution of very high susceptibility zones (%)	Key associations with geographic features	Area percentage of very high and high zones (%)	Landslide points in very high and high-risk zones (%)
IV-SVM	11.15	Distance from river	24.77	88.15
IV-LR	5.85	Distance from road	18.20	88.89
IV-RF	10.32	Distance from road	21.86	84.44
IV-ANN	12.14	Distance from river	21.32	83.70

Figure 7

Area percentage of different susceptibility classes of IV; WOE; IV-LR; and IV-ANN.

3.3 Uncertainty analysis of modeling results

Due to the highly diverse factors influencing the occurrence of landslides and the unknown accuracy of the zoning results [42], this study utilized common model evaluation metrics such as precision, recall, accuracy, and F1 to assess the accuracy of the zoning results and the generalization ability of the model. The evaluation indicators are shown in Table 7 [43]. TP represents the number of positive samples that are correctly classified by the evaluation model, that is, the true examples. TN represents the number of negative samples that are wrongly classified by the evaluation model, that is, true counterexamples. FP represents the number of positive samples that are misclassified by the evaluation model, that is, false positive examples. FN represents the number of negative samples that are misclassified by the evaluation model, that is, false counterexamples. After calculation, the accuracy of the IV-SVM, IV-LR, V-RF and IV-ANN models on the test set are 76.54%, 77.78%, 82.72% and 76.54% respectively. The difference in prediction accuracy rates of the models is not significant, but they are all greater than 70%. Among them, the IV-RF model has the best prediction accuracy, and all can provide effective evaluations for the susceptibility of landslides in Minfeng County.

The receiver operating characteristics (ROC) [46] curve is an effective method for evaluating the accuracy of landslides, and its longitudinal axis is sensitivity, horizontal axis is 1-specificity, and area under curve (AUC) [47], the larger the value, the better the model evaluation effect. Figure 8 shows the results of analyzing the ROC curve.

Table 7

Model evaluation index table

	Accuracy	Precision	Recall	F1_score	AUC
IV-SVM	0.7654	0.7692	0.7500	0.7595	0.8537
IV-LR	0.7778	0.7619	0.8000	0.7805	0.8537
IV-RF	0.8272	0.8250	0.8250	0.8250	0.8945
IV-ANN	0.7654	0.7561	0.7750	0.7654	0.8768
Calculation formula	( TP + TN ) ( TP + TN + FP + FN )	TP ( TP + FP )	( TP ) ( TP + FN )	2 × ( precision × recall ) ( precision + recall )	ROC

Figure 8

ROC plot.

It can be seen from Figure 8 that the AUC values of the IV-SVM, IV-LR, IV-RF, and IV-ANN models are 0.854, 0.854, 0.895, and 0.877, respectively. The ROC curves of the four coupled models were compared. The AUC values indicated that the modeling accuracy of the four models was not much different, all above 0.8. The prediction results of the models were good, among which the prediction results of the IV-RF models were the best.

4 Discussion

In this study, by means of the coupling of IV information quantity values and machine learning, the accuracy of the model results was compared, and the most suitable method for evaluating the susceptibility of landslide disasters in the complex arid mountainous areas of Xinjiang was selected. The evaluation indicators of each model were suitable for the susceptibility evaluation of this region. However, the accuracy of the IV-RF model was 0.8272, and the AUC area of the ROC curve was 0.8945. It is the best in comparison with other coupled models. In this study, compared with previous literature, there are also certain improvements and innovations. For example, compared with the Abhik Saha [46] study, the method of coupling the statistical model and the machine learning model proposed in this study is more suitable for the evaluation of complex terrain areas. Compared with the research of Gu et al. [47], although there is a combination of hybrid models, the cases applied in arid and complex mountainous areas are still relatively few. Compared with the studies of Yesilnacar and Topal [48], in the model validation and evaluation, combining multiple evaluation indicators, cross-validation, and other methods makes the model evaluation more objective.

Although this research has achieved certain results, there are still some limitations. The 135 landslide data from 2020 to 2023 used in this study are mainly based on remote sensing interpretation and limited field investigations. Due to the complex terrain and inconvenient transportation in the high-altitude uninhabited area in the south and the desert edge area in the north of the study area, there may be omissions, which may affect the accuracy of the model. In addition, small rock avalanches may not be captured by remote sensing images, resulting in insufficient representation of samples in the low and moderately sensitive areas of the inventory, which in turn has a slight impact on the calculation of model factor weights. The accuracy of 30 m resolution DEM data is limited when extracting micro-terrain features, which may lead to errors in classification factors such as slope and slope direction. The geological map (1:500,000) has a small scale, and the local fault zones and lithological boundaries are not depicted precisely enough, which may affect the accuracy of factors such as the “distance from the fault.” Auxiliary data, such as NDVI, are all average values from 2020 to 2023 and cannot fully reflect the dynamic changes in vegetation coverage before and after a single rainstorm or extreme weather event. Some landslides may be more related to short-term vegetation destruction. This mismatch on the time scale may weaken the correlation analysis between the NDVI factor and landslides. Furthermore, the construction of the model is based on existing data and methods, which may not be fully applicable to the complex and variable environmental conditions in the future, such as an increase in extreme climate events or significant changes in the intensity of human activities.

In view of the limitations of this study, future research can be carried out on the following aspects. First, further improve the landslide list by integrating more field investigations and high-precision remote sensing technology to enhance the accuracy and completeness of the data. Second, conduct in-depth research on other potential influencing factors and explore how to incorporate them into the landslide susceptibility assessment model. For instance, carry out research on the relationship between ground motion parameters and landslides, analyze the landslide triggering mechanisms under different rainfall patterns, and thereby construct a more comprehensive assessment system. Third, it is necessary to enhance the research on the dynamic adaptability of models, taking into account the impact of climate change and dynamic changes in human activities on landslides, and develop models that can be updated and predicted in real time to improve the early warning capacity for landslide disasters and the scientific nature of response measures. Fourth, integrate the assessment of landslide susceptibility with risk assessment, not only focusing on the possibility of landslides occurring but also considering the potential losses they may cause, providing more practical basis for disaster management and decision-making.

5 Conclusion

The susceptibility of landslides in Minfeng County, Xinjiang, was evaluated by coupling four information quantity methods with machine learning models. The conclusions are as follows:

The main influencing factors of the high and extremely high landslide susceptibility areas in Minfeng County are the distance from the water system and the distance from the road. The susceptibility mainly occurs in the area about 400 m away from the water system, distributed in the central part of Minfeng County, which is also an area with dense human engineering activities. However, the overall high-incidence area accounts for a relatively small proportion, and it is mainly low landslide susceptibility areas. The prediction results of different coupling models in the study area have certain similarities.
The results of the IV-SVM, IV-RF, IV-LR, and IV-ANN models show that as the susceptibility level of landslides increases, the number of landslide points within the interval gradually increases. More than half of the landslide points in each information coupling model are located in the high-very high susceptibility area, while there are very few landslide points in the low-very low susceptibility area. This is consistent with the actual landslide disaster situation, indicating that each model can effectively evaluate the susceptibility of landslide disasters in Minfeng County.
The verification accuracies of the IV-SVM, IV-RF, IV-LR, and IV-ANN models were 76.54, 77.78, 82.72, and 76.54%, respectively, and the AUC values were 0.854, 0.854, 0.895, and 0.877, respectively. Combining the two verification results, it was concluded that the IV-RF model has the best effect on the zoning of landslide susceptibility in Minfeng County, with the highest classification accuracy. It can well reflect the differences in landslide susceptibility among different grids.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (62101395).

Funding information: Authors state no funding involved.
Author contributions: Yangyang Jiao and Daozhu Xu wrote the main manuscript text. Qiang Wang and Lei Wang prepared Figures 1–4. All authors reviewed the manuscript.
Conflict of interest: The authors declare no conflicts of interest.
Data availability statement: All data generated or analyzed during this study are included in this published article. The data underlying this article will be shared on reasonable request to the corresponding author.

References

[1] Bathrellos GD, Koukouvelas IK, Skilodimou HD, Nikolakopoulos KG, Vgenopoulos A-L. Landslide causative factors evaluation using GIS in the tectonically active Glafkos River area, northwestern Peloponnese, Greece. Geomorphology. 2024;461(Sep.15):1.1–8.10.1016/j.geomorph.2024.109285Suche in Google Scholar

[2] Karpouza M, Skilodimou HD, Kaviris G, Zymvragakis A, Antonarakou A, Bathrellos GD. Escape routes and safe points in natural hazards. A case study for soil. Eng Geol. 2024;340:107683.10.1016/j.enggeo.2024.107683Suche in Google Scholar

[3] Ladel L, Mastere M, Kader S, Spalević V, Dudic B. Machine learning methods for landslide mapping studies: A comparative study of SVM and RF algorithms in the Oued Aoulai watershed (Morocco). Open Geosci. 2025;17(1):20220740.10.1515/geo-2022-0740Suche in Google Scholar

[4] Ullah S, Hanif M, Radwan AE, Luo C, Ur Rehman N, Ahmad S, et al. Depositional and diagenetic modeling of the Margala Hill Limestone, Hazara area (Pakistan): Implications for reservoir characterization using outcrop analogues. Geoenergy Sci Eng. 2023;224:211584.10.1016/j.geoen.2023.211584Suche in Google Scholar

[5] Saha A, Villuri VGK, Bhardwaj A. Development and assessment of a novel hybrid machine learning-based landslide susceptibility mapping model in the Darjeeling Himalayas. Stoch Environ Res Risk Assess. 2023;3:1–24.10.1007/s00477-023-02528-8Suche in Google Scholar

[6] Youssef AM, El‑Haddad BA, Skilodimou HD, Bathrellos GD, Golkar F, Pourghasemi HR. Landslide susceptibility, ensemble machine learning, and accuracy methods in the southern Sinai Peninsula, Egypt: Assessment and Mapping. Nat Hazards. 2024;120:14227–58.10.1007/s11069-024-06769-wSuche in Google Scholar

[7] Gu T, Duan P, Wang M, Li J, Zhang Y. Effects of non-landslide sampling strategies on machine learning models in landslide susceptibility mapping. Sci Rep. 2024;14:7201.10.1038/s41598-024-57964-5Suche in Google Scholar PubMed PubMed Central

[8] Saha A, Tripathi L, Villuri VGK, Bhardwaj A. Exploring machine learning and statistical approach techniques for landslide susceptibility mapping in Siwalik Himalayan Region using geospatial technology. Environ Sci Pollut Res. 2024;31(7):10443–59.10.1007/s11356-023-31670-7Suche in Google Scholar PubMed

[9] Abdelkader MM, Csámer Á. Comparative assessment of machine learning models for landslide susceptibility mapping: a focus on validation and accuracy. Nat Hazards. 2025;121:1–23.10.1007/s11069-025-07197-0Suche in Google Scholar

[10] Sarfraz Y, Basharat M, Riaz MT, Akram MS, Xu C, Ahmed KS, et al. Application of statistical and machine learning techniques for landslide susceptibility mapping in the Himalayan road corridors. Open Geosci. 2022;14(1):1606–35.10.1515/geo-2022-0424Suche in Google Scholar

[11] Alabert FG, Massonnat GJ. Heterogeneity in a complex turbiditic reservoir: stochastic modelling of facies and petrophysical variability. SPE Annual Technical Conference and Exhibition, New Orleans, Louisiana; 1990.10.2118/20604-MSSuche in Google Scholar

[12] Al-Mudhafar WJ. Multiple-point geostatistical lithofacies simulation of fluvial sand-rich depositional environment: a case study from Zubair Formation/South Rumaila Oil Field. SPE Reserv Eval Eng. 2017;21(1):39–53.10.2118/187949-PASuche in Google Scholar

[13] Teng F, Mao Y, Li Y, Qian S, Nanehkaran YA. Comparative models of support-vector machine, multilayer perceptron, and decision tree ‎predication approaches for landslide ‎susceptibility analysis. Open Geosci. 2024;16(1):20220642.10.1515/geo-2022-0642Suche in Google Scholar

[14] Radwan AE. Modeling the depositional environment of the sandstone reservoir in the Middle Miocene Sidri Member, Badri Field, Gulf of Suez Basin, Egypt: Integration of gamma-ray log patterns and petrographic characteristics of lithology. Nat Resour Res. 2021;30:431–49.10.1007/s11053-020-09757-6Suche in Google Scholar

[15] Seifert D, Jensen JL. Using sequential indicator simulation as a tool in reservoir description: issues and uncertainties. Math Geol. 1999;31(5):527–50.10.1023/A:1007563907124Suche in Google Scholar

[16] Aleksova B, Milevski I, Mijalov R, Marković SB, Cvetković VM, Lukić T. Assessing risk-prone areas in the Kratovska Reka catchment (North Macedonia) by integrating advanced geospatial analytics and flash flood potential index. Open Geosci. 2024;16(1):20220684.10.1515/geo-2022-0684Suche in Google Scholar

[17] Zhang T. Incorporating geological conceptual models and interpretations into reservoir modeling using multiple-point geostatistics. Earth Sci Front. 2008;15(1):26–35.10.1016/S1872-5791(08)60016-0Suche in Google Scholar

[18] Liu Y, Harding A, Abriel W, Strebelle S. Multiple-point simulation integrating wells, three-dimensional seismic data, and geology. AAPG Bull. 2004;88(7):905–21.10.1306/02170403078Suche in Google Scholar

[19] Journel AG. Geostatistics for reservoir characterization. SPE Annual Technical Conference and Exhibition, New Orleans, Louisiana; 1990.10.2118/20750-MSSuche in Google Scholar

[20] Li P, Zhao Y, Sufian M, Deifalla AF. Scientometric analysis of flood forecasting for Asia region and discussion on machine learning methods. Open Geosci. 2023;15(1):20220475.10.1515/geo-2022-0475Suche in Google Scholar

[21] Al-Mudhafar WJ. Bayesian kriging for reproducing reservoir heterogeneity in a tidal depositional environment of a sandstone formation: a case study. J Appl Geophysics. 2018;160(1):84–102.10.1016/j.jappgeo.2018.11.007Suche in Google Scholar

[22] Doyen PM, Den Boer LD, Pillet WR. Seismic porosity mapping in the Ekofisk field using a new form of collocated cokriging. SPE Annual Technical Conference and Exhibition, Denver, Colorado; 1996.10.2118/36498-MSSuche in Google Scholar

[23] Al-Mudhafar WJ, Vo Thanh H, Wood DA, Min B. Stochastic lithofacies and petrophysical property modeling for fast history matching in heterogeneous clastic reservoir applications. Sci Rep. 2024;14:22.10.1038/s41598-023-50853-3Suche in Google Scholar PubMed PubMed Central

[24] Journel AG, Alabert FG. New method for reservoir mapping. J Pet Technol. 1990;42(2):212–8.10.2118/18324-PASuche in Google Scholar

[25] Al-Mudhafar WJ. Geostatistical simulation of facies and petrophysical properties for heterogeneity modeling in a tidal depositional environment: a case study from upper shale member in A Southern Iraqi Oil Field. Unconventional Resources Technology Conference, Houston, TX; 2021.10.15530/urtec-2021-5551Suche in Google Scholar

[26] Xu W, Tran TT, Srivastava RM, Journel AG. Integrating seismic data in reservoir modeling: the collocated cokriging alternative. SPE Annual Technical Conference and Exhibition, Washington, DC; 1992.10.2523/24742-MSSuche in Google Scholar

[27] Petley D. Global patterns of loss of life from landslides. Geology. 2012;40(10):927–30.10.1130/G33217.1Suche in Google Scholar

[28] Ye P, Yu B, Chen W, Liu K, Ye L. Rainfall-induced landslide susceptibility mapping using machine learning algorithms and comparison of their performance in Hilly area of Fujian Province, China. Nat Hazards. 2022;113(2):965–95.10.1007/s11069-022-05332-9Suche in Google Scholar

[29] Yesilnacar E, Topal T. Landslide susceptibility mapping: a comparison of logistic regression and neural networks methods in a medium scale study, Hendek region (Turkey). Eng Geol. 2005;79(3–4):251–66.Suche in Google Scholar

[30] Oliveira SC, Zêzere JL, Garcia RAC, Pereira S, Vaz T, Melo R. Landslide susceptibility assessment using different rainfall event-based landslide inventories: advantages and limitations. Nat Hazards. 2024;120:9361–99.10.1007/s11069-024-06691-1Suche in Google Scholar

[31] Singh A, Dhiman N, Niraj KC, Shukla DP. Ensembled transfer learning approach for error reduction in landslide susceptibility mapping of the data scarce region. Sci Rep. 2024;14:29060.10.1038/s41598-024-76541-4Suche in Google Scholar PubMed PubMed Central

[32] Ajin RS, Segoni S, Fanti R. Optimization of SVR and CatBoost models using metaheuristic algorithms to assess landslide susceptibility. Sci Rep. 2024;14:24851.10.1038/s41598-024-72663-xSuche in Google Scholar PubMed PubMed Central

[33] Francis DM, Bryson LS. Coupled landslide analyses through dynamic susceptibility and forecastable hazard analysis. Nat Hazards. 2025;121:2971–99.10.1007/s11069-024-06908-3Suche in Google Scholar

[34] Edrich AK, Yildiz A, Roscher R, Bast A, Graf F, Kowalski J. A modular framework for FAIR shallow landslide susceptibility mapping based on machine learning. Nat Hazards. 2024;120:8953–82.10.1007/s11069-024-06563-8Suche in Google Scholar

[35] Bornaetxea T, Remondo J, Bonachea J, Valenzuela P. Exploring available landslide inventories for susceptibility analysis in Gipuzkoa province (Spain). Nat Hazards. 2023;118:2513–42.10.1007/s11069-023-06103-wSuche in Google Scholar

[36] Sun D, Wen H, Wang D, Xu J. A random forest model of landslide susceptibility mapping based on hyperparameter optimization using Bayes algorithm. Geomorphology. 2020;362:107201.10.1016/j.geomorph.2020.107201Suche in Google Scholar

[37] Sun D, Gu Q, Wen H, Xu J, Zhang Y, Shi S, et al. Assessment of landslide susceptibility along mountain highways based on different machine learning algorithms and mapping units by hybrid factors screening and sample optimization. Gondwana Res. 2023;123:89–106.10.1016/j.gr.2022.07.013Suche in Google Scholar

[38] Kikuchi T, Sakita K, Nishiyama S, Takahashi K. Landslide susceptibility mapping using automatically constructed CNN architectures with pre-slide topographic DEM of deep-seated catastrophic landslides caused by Typhoon Talas. Nat Hazards. 2023;117:339–64.10.1007/s11069-023-05862-wSuche in Google Scholar

[39] Khan I, Kainthola A, Bahuguna H, Pandey VH, Asgher MS, Bhardwaj A, et al. Integrating FR, MFR and IV models for landslide susceptibility zonation mapping in Joshimath Watershed, Uttarakhand, India. Iran J Sci Technol Trans Civ Eng. 2024;25(1):5207–31.10.1007/s40996-024-01708-ySuche in Google Scholar

[40] Dyer AS, Mark-Moser M, Duran R, Bauer JR. Offshore application of landslide susceptibility mapping using gradient-boosted decision trees: a Gulf of Mexico case study. Nat Hazards. 2024;120:6223–44.10.1007/s11069-024-06492-6Suche in Google Scholar

[41] Harvey EL, Kincey ME, Rosser NJ, Gadtaula A, Collins E, Densmore AL, et al. Review of landslide inventories for Nepal between 2010 and 2021 reveals data gaps in global landslide hotspot. Nat Hazards. 2025;121:5075–101.10.1007/s11069-024-07013-1Suche in Google Scholar

[42] Sun D, Ding Y, Zhang J, Wen H, Wang Y, Xu J, et al. Essential insights into decision mechanism of landslide susceptibility mapping based on different machine learning models. Geocarto Int. 2022;1(29):1–29.10.1080/10106049.2022.2146763Suche in Google Scholar

[43] Zhang L, Zeng H, Ding Y, Hu H, Chen L, Zhang J, et al. Geo-environment-aware adversarial transfer learning method for landslide susceptibility evaluation of complex mountainous areas. Trans GIS. 2023;27(5):1418–40.10.1111/tgis.13080Suche in Google Scholar

[44] Durmaz M, Hürlimann M, Huvaj N, Medina V. Comparison of different hydrological and stability assumptions for physically-based modeling of shallow landslides. Eng Geol. 2023;323:107237.10.1016/j.enggeo.2023.107237Suche in Google Scholar

[45] Su C, Wang B, Lv Y, Zhang M, Peng D, Bate B, et al. Improved landslide susceptibility mapping using unsupervised and supervised collaborative machine learning models. Georisk: Assess Manag Risk Eng Syst Geohazards. 2023;17(2):387–405.10.1080/17499518.2022.2088802Suche in Google Scholar

[46] Saha A, Villuri VGK, Bhardwaj A. Development and assessment of GIS-based landslide susceptibility mapping models using ANN, Fuzzy-AHP, and MCDA in Darjeeling Himalayas, West Bengal, India. Land. 2022;11(10):1711.10.3390/land11101711Suche in Google Scholar

[47] Gu T, Li J, Wang M, Duan P, Zhang Y, Cheng L. Study on landslide susceptibility mapping with different factor screening methods and random forest models. PLoS One. 2023;18(10):e0292897.10.1371/journal.pone.0292897Suche in Google Scholar PubMed PubMed Central

[48] Yesilnacar E, Topal T. Landslide susceptibility mapping: a comparison of logistic regression and neural networks methods in a medium scale study, Hendek region (Turkey). Eng Geo. 2005;79(3–4):251–66.10.1016/j.enggeo.2005.02.002Suche in Google Scholar

Received: 2025-06-13

Revised: 2025-08-13

Accepted: 2025-09-24

Published Online: 2025-11-05

This work is licensed under the Creative Commons Attribution 4.0 International License.

Artikel in diesem Heft

https://doi.org/10.1515/geo-2025-0906

Schlagwörter für diesen Artikel

landslide risk assessment; hybrid modeling, information value model; random forest; logistic regression; artificial neural network; support vector machine; arid mountainous regions; Xinjiang

Creative Commons

BY 4.0