Application of a Hybrid Model in Landslide Susceptibility Evaluation of the Western Tibet Plateau

Abstract

The evaluation of landslide susceptibility plays a crucial role in preventing the risks associated with landslides and debris flows, providing valuable insights for the effective prevention and mitigation of geological hazards. However, there is limited research on high-altitude areas. Therefore, this study chose the western Tibetan Plateau as the study area, a representative area known for its susceptibility to landslides and high attitudes. In this study, seven factors were identified based on research objectives. Information value (IVM), weight of evidence (WOE), information value logistic regression (IVM-LR), weight of evidence logistic regression (WOE-LR), information value multi-layer perceptron (IVM-MLP) and weight of evidence multi-layer perceptron (WOE-MLP) were selected and compared for landslide susceptibility. The percentage of disaster area included in each risk level, the AUC value and the ROC curve were used to evaluate the accuracy of the results. The ROC curves of the results were close to the upper–left corner and the AUC values exceeded 0.85, an indication that all results were highly accurate. Moreover, the percentage of disaster area included for each risk showed an upward trend regarding susceptibility. The results indicated that the hybrid model exhibited superior performance in assessing landslide susceptibility at high altitudes. Overall, the results showed great significance regarding disaster prevention and mitigation measures of local governments.

1. Introduction

Landslides are a prevalent geological hazard, causing significant damage and loss of life on a widespread basis [1,2,3,4]. Consequently, global research efforts are focused on landslide hazard prevention [5,6,7]. Geographic Information System (GIS) technology has provided an effective means of studying landslide susceptibility [8,9,10,11]. Various methods have been developed for assessing landslide susceptibility and can be divided into traditional and machine learning models [12,13,14,15,16,17,18,19,20,21,22,23,24]. For instance, Kayastha (2013) employed an analytic hierarchy process model to analyze landslide susceptibility in the vicinity of Beihai City, China, and the Tinau watershed in Western Nepal [25]. Ozdemir (2013) investigated the comparative efficacy of logistic regression (LR) and evidence methods using the Sultan Mountains in SW Turkey [26]. Sharma (2014) utilized the information value model (IVM) to determine landslide susceptibility in the Sikkim Himalayas, India [27]. Zhao (2021) conducted a study employing a machine learning support vector machine, a certainty factor, and random forest models in Ningqiang County, Shanxi Province, China [28]. Other models include adaptive network-based fuzzy inference models, the weight of evidence model (WoE), multi-layer perceptron (MLP), support vector machines (SVM), and decision trees [29,30,31,32,33,34,35]. Many studies have used more than two models to determine difference in the results. However, these results show that the learning ability of the traditional models is relatively weak, and machine learning models need sufficient and balanced data. There is no best single or mixed model that can be used in all areas. High-altitude areas have less interference from human factors and provide a good demonstration area for the comparative analysis of different models. At the same time, the roads in these areas are relatively simple; however, disasters may cause great damage to local traffic, and so they have extremely important practical research significance. In addition, limited research has been conducted in high-altitude areas of the Tibetan Plateau.

Based on this context, the present study selected Zhipuqi, which has the lowest altitude (3869 m) in the region, as the study area (Figure 1). Landslides represent one of the most significant geological hazards in this area. Analyzing landslide susceptibility mapping has considerable practical implications for planning and disaster prevention, significantly enhancing the safety of lives and property. To assess the feasibility of geological hazard analysis in high-altitude areas and to compare model differences, we employed the IVM [36], the weight of evidence (WoE) method [37,38,39,40], the LR model [41,42,43,44,45] and the multi-layer perceptron (MLP) method [46,47,48]. Of these models, the IVM and WoE can more quickly obtain a stable weight of each low-cost factor, and the LR and MLP can provide a higher learning ability for mixed models.

2. Materials and Methods

2.1. Study Area

The study area is located in the Zada basin, the west of the Tibetan Plateau; it belongs to the North Himalayan Tethys sedimentary zone. It lies in the junction zone of Himalayan land mass and Brahmaputra River. Between 78°15′–79°00′ E and 32°20′–32°40′ N, the total area covers 2600 km² (Figure 1). It lies in a typical alpine canyon landform, an elevation ranging from 3869 to 6357 m, with most peaks more than 5000 m above sea level. It belongs to the typical alpine canyon landform and it is mainly composed of hilly landforms and deep canyon forming bytectonic faults. The only river is Pali river, which flows through the urban area from NW to SE with elevation between 3000 and 3800 m. This area has cold and dry climatic conditions. The mean annual temperature is between 0 and 3 °C with an annual precipitation of 200 mm. The maximum summer temperatures reach above 21 °C, whereas winter temperatures can be lower than −41 °C. The rock strata in the study area are mainly in the Holocene Series, Pleistocene Series, Selong Group and Cailiqun Group, and the rock types are mainly soil, gray and purple gray mesobioclastic micrite, carbonate rock (limestone, oolitic limestone, biological limestone) mainly composed of sand-shale assemblages. The rock structure is mainly block structure. Due to the special geographical and geological conditions of the region, it is of great importance to carry out disaster risk assessments, and it is also a good area for conducting high-altitude research.

2.2. Data Source and Methodology

Landslide inventories of this area are important for producing landslide susceptibility maps [49]. According to a detailed landslide survey, 151 landslides were collected (Figure 1). All landslides were selected for statistical landslide susceptibility analysis [50]. The major data of this study include a topographic and geological map of 1:200,000, a 30 m resolution DEM, Landsat-8 satellite remote sensing images, and existing reports and field survey data of hazards (

Table 1. Sources of thematic factors.

Base Map	Thematic Factor	Source
DEM	Slope	ASTER DEM (30 m)
	Aspect
	Elevation
Geological map	Lithology	China National Archive 1:200,000
	Distance to faults
Geographic map	Distance to roads	National Geomatic Center of China
	Distance to rivers

). This study relied on IVM and the WoE value to calculate the risk weight of factors, and the LR and MLP models were used to calculate the weight coefficient. Finally, the receiver operating characteristic (ROC) curve was used to analyze the accuracy of model predictions. The methodology of the study area is shown as a flowchart (Figure 2), which shows the key steps to obtaining the landslide susceptibility map of the area.

2.3. Conditioning Factors

Based on the literature, effectiveness, availability of data and relevance with respect to landslide occurrence, seven factors were chosen to analyze the landslide: slope, aspect, elevation, lithology, distance to faults, distance to roads, and distance to rivers to produce the landslide susceptibility map.

2.3.1. DEM and Derivatives

Many common factors can be found using DEM, such as slope, slope aspect, and elevation [51]. In this area, the slope generally ranged from 0° to 60°, and was grouped into 7 classes: 0–10°, 10–20°, 20–30°, 30–40°, 40–50°, 50–60°, >60° (Figure 3a). The slope aspect affecting sunlight and wind was classified into nine categories: flat, north, northeast, east, southeast, south, southwest, west, northwest (Figure 3b). The elevation was between 3869 and 6357 m, which was divided into eight classes with intervals of 300 m (Figure 3c).

2.3.2. Lithology

Lithology has a very important effect on the occurrence of landslides because different lithologies can produce different influences during landslide occurrence [52]. The lithology of the area primarily consists of sand slate, limestone, breccia, and quaternary sediments. Influenced by the intensity of rock weathering and soil formation, according to the hardness, it can be divided into the following categories: hard rock, harder rock, softer rock, soft rock and extremely soft rock according to engineering geology (Figure 3d).

2.3.3. Distance to Faults

Faults cause fragmentation of the surrounding rocks and provide a source of material and power for the generation of landslides [53]. According to the distance to the faults, the work was divided into eight grades at the intervals of 300 m: 0–300, 300–600, 600–900, 900–1200, 1200–1500, 1500–1800, 1800–2100, >2100 m (Figure 3e).

2.3.4. Distance to Roads

Road vicinity is a human-made factor closely related to landslide formation [54]. The road construction process can destabilize the slopes on both sides of the road and strengthen the degree of rock fragmentation. The road data obtained were gridded, and a buffer zone was divided into eight classes established at intervals of 300 m from the road (Figure 3f).

2.3.5. Distance to River

Rivers can strengthen the erosion of surrounding rocks and invade rock cracks, thereby providing downward power for the formation of landslides [45]. It was divided into eight grades at 300 m intervals (Figure 3g).

2.4. Multicollinearity Diagnosis

IBM SPSS Statistics 21 was used to conduct multi-collinear diagnosis analysis. The results are presented in

Table 2. Conditioning factor categories.

Factor	TOL	VIF
Slope degree	0.937	1.077
Slope aspect	0.967	1.034
Elevation	0.929	1.077
Lithology	0.695	1.439
Distance to faults	0.867	1.154
Distance to roads	0.832	1.203
Distance to rivers	0.943	1.06

and

Table 3. The results of collinearity diagnostics.

Dimensionality	Eigenvalue	Condition Index	Variance Proportion
			Slope	Aspect	Elevation	Lithology	Distance to Faults	Distance to Roads	Distance to Rivers	Slope Degree
1	1.918	1	0	0.03	0.03	0.06	0.12	0.06	0.08	0.04
2	1.13	1.303	0.41	0.14	0	0	0	0.19	0.05	0.02
3	1.056	1.348	0.15	0.17	0.08	0.05	0.03	0.07	0.16	0.15
4	0.955	1.417	0	0	0.67	0.02	0	0.04	0.03	0.22
5	0.896	1.463	0.07	0.36	0.03	0.53	0.01	0.01	0.01	0.01
6	0.846	1.506	0.06	0.12	0.12	0.29	0.01	0.03	0	0.47
7	0.697	1.659	0.3	0.11	0.05	0	0	0.32	0.4	0.08
8	0.503	1.953	0	0.06	0.01	0.05	0.83	0.27	0.28	0

. It can be seen from the table that among the selected factors, the maximum variance inflation factor (VIF) of lithology was 1.439 < 10, the minimum tolerance (Tol) value was 0.695 > 0.1, the minimum eigenvalue was 0.503 > 0 and the maximum condition Idex was 0.503 < 10, indicating that there was no serious collinearity problem. Therefore, these seven selected factors could be used for landslide sensitivity analysis [28].

2.5. Landslide Susceptibility Modeling

2.5.1. Information Value Model (IVM)

The IVM is one of the most commonly used methods for geological disaster risk assessments [36,55]. It mainly converts the collected real geological disaster-related impact factor data into the information quantity required for calculation and obtains the total information quantity by superposing the information quantity of each factor. The specific theoretical and formula calculation processes of the landslide susceptibility index (LSI) are as follows:

$$\mathrm{L}\mathrm{S}\mathrm{I}={\sum }_{i=1}^{n}log2\frac{Bi/B}{Si/S},$$

(1)

where S refers to the total area, B represents the area of grids containing geological hazards, and Si and Bi represent the total area or the area grids containing geological hazards of i.

2.5.2. Weight of Evidence Method (WoE)

Bonham–Carter (1988) first used the WoE to carry out mineral potential assessment [56]. Since then, the WoE has been applied in many fields [57,58,59,60,61]. The WoE is based on Bayesian statistical analysis, and the specific process of the model is as follows:

First, the total grid number of the workspace is preset as N[Z]. There is only one geological disaster point in each grid range. The initial probability of a geological disaster occurring in the working area can be obtained as N[L]/N[Z].

Second, the proportion of different factors at different stages in the probability of occurrence of geological disasters can be obtained by presetting W_i⁺ and W_i⁻. W_i⁺ and W_i⁻, respectively, represent the positive and negative effects of factors at this stage on the likelihood of occurrence of geological disasters:

$${\mathrm{W}}_{\mathrm{i}}^{+}=\ln \frac{\mathrm{P}\left\{\mathrm{N}\right[\mathrm{X}\mathrm{i}\left]\right|\mathrm{N}\left[\mathrm{L}\right]\}}{\mathrm{P}\left\{\mathrm{N}\right[\mathrm{X}\mathrm{i}\left]\right|\mathrm{N}\left[\overline{\mathrm{L}}\right]\}}=\mathrm{l}\mathrm{n}\frac{\mathrm{N}(\mathrm{X}\mathrm{i}\cap \mathrm{L})/\mathrm{N}\left(\mathrm{L}\right)}{\mathrm{N}(\mathrm{X}\mathrm{i}\cap \overline{\mathrm{L}})/\mathrm{N}\left(\overline{\mathrm{L}}\right)},$$

(2)

$${\mathrm{W}}_{\mathrm{i}}^{-}=\ln \frac{\mathrm{P}\left\{\mathrm{N}\right[\overline{\mathrm{X}\mathrm{i}}\left]\right|\mathrm{N}\left[\mathrm{L}\right]\}}{\mathrm{P}\left\{\mathrm{N}\right[\overline{\mathrm{X}\mathrm{i}}\left]\right|\mathrm{N}\left[\overline{\mathrm{L}}\right]\}}=\mathrm{l}\mathrm{n}\frac{\mathrm{N}(\overline{\mathrm{X}\mathrm{i}}\cap \mathrm{L})/\mathrm{N}\left(\mathrm{L}\right)}{\mathrm{N}(\overline{\mathrm{X}\mathrm{i}}\cap \overline{\mathrm{L}})/\mathrm{N}\left(\overline{\mathrm{L}}\right)}.$$

(3)

In the above formulae, the different stages of the different geological disaster impact factors are noted as Xi. The number of grids contained in this stage is written as N[Xi], and the grid number with geological hazard points is N[L]. In the instance where no geological hazard is present, they are denoted as N[$\overline{Xi}$] and N[$\overline{L}$], respectively.

N(Xi∩L) is the number of grids containing geological hazards of Xi. In the case of everything else, the converse condition is Xi∩$\overline{L}$.

Then, the contrasting weights can be obtained by addition and subtraction.

$${\mathrm{W}}^{+}={\mathrm{W}}_i^{+}+{\mathrm{W}}_i^{-},$$

(4)

$${\mathrm{W}}^{-}={\mathrm{W}}_i^{+}-{\mathrm{W}}_i^{-}.$$

(5)

Finally, the LSI can be computed by summing up the effects, allowing for the delineation of an LSZ map.

$$\mathrm{L}\mathrm{S}\mathrm{I}=\sum \mathrm{W}.$$

(6)

Finally, two Landslide susceptibility mappings can be acquired by the WoE model. W⁺ can amplify the subtle differences in the LSI by the positive and negative effects of factors adding together. W⁻ can find the dominant positive and negative factors is dominant.

2.5.3. Logistic Regression (LR)

The LR model is mainly used to infer the relationship between a group of relatively independent geological environmental factors and the occurrence of geological disasters. In other words, the model seeks to determine whether geological disasters can occur under such conditions, and if yes, the probability of their occurrence [62,63,64]. In this model, the probability of geological disasters is defined between 0 and 1, and the probability of hazard distribution in the working area is then calculated [34]. Compared with the analytic hierarchy process, the geological environment factors selected by this model can be either discrete or continuous, which can more reasonably fit the probability of landslide occurrence [65]. The specific calculation process is as follows:

$$\mathrm{A}=\frac{\exp \left(\mathrm{M}\right)}{1+\exp \left(\mathrm{M}\right)}.$$

(7)

A indicates the probability of a landslide occurrence from 0 to 1. M is a linear combination, and can be expressed as

$$\mathrm{A}=\frac{\exp \left(\mathrm{M}\right)}{1+\exp \left(\mathrm{M}\right)}.$$

(8)

X1, X2, …, and Xn denote the variables. N1, N2, …, and Nn are the slope coefficients. N0 refers to the intercept.

2.5.4. Multi-Layer Perceptron (MLP)

Multi-layer perceptron (MLP) is one of the most frequently used artificial neural network methods with three layers combining input, hidden and output layers [66]. It can find a non-linear relationship between a lot of independent parameters.

The associative matrix of processing objects is constructed by joining relations between input layers and hidden layers. The connection between hidden layer and output layer elements forms the decision matrix of the processing object. Through training, the network can form an orderly and stable structure with a decision-making ability. The back-propagation (BP) algorithm is widely used in an MLP model. The BP algorithm consists of forward transmission of information and back propagation of error. Forward transmission. In the process of seeding, the input information is calculated from the input layer to the output layer through the hidden layer, and the state of neurons in each layer only affects the state of the next layer of neurons. If the desired output is not obtained in the output layer, the error variation of the output layer is calculated, and then reverse propagation is carried out. Through the network, the error information is reversed back along the original link path to modify the weight of neurons in each layer until the desired goal is reached.

3. Results

3.1. Landslide Susceptibility Mapping by the IVM Model

The relationships between the factors calculated by the information model and disasters are presented in

Table 4. Weight of factors for the IVM, W⁺, W⁻.

Factors	Class	Class Pixel Counts	Class Pixel Counts %	Landslide Pixel Counts	Landslide Pixel Counts %	${\mathbf{W}}_{\mathit{i}}^{+}$	${\mathbf{W}}_{\mathit{i}}^{-}$	W+	W−	IVM
Slope	0–10	121,295	12.29	3699	6.97	−0.592	0.062	−0.529	−0.654	−0.819
	10–20	255,220	25.85	14,338	27.00	0.046	−0.017	0.030	0.063	0.063
	20–30	329,166	33.34	21,463	40.42	0.205	−0.118	0.086	0.323	0.278
	30–40	216,075	21.89	11,389	21.45	−0.021	0.006	−0.015	−0.027	−0.029
	40–50	58,428	5.92	2084	3.92	−0.430	0.022	−0.408	−0.452	−0.593
	50–60	6637	0.67	128	0.24	−1.061	0.005	−1.057	−1.066	−1.480
	>60	487	0.05	0	0.00	0.000	0.001	0.000	0.000	0.000
Aspect	Flat	2630	0.27	131	0.25	−0.081	0.000	−0.081	−0.081	−0.111
	North	143,044	14.49	8196	15.43	0.067	−0.012	0.055	0.079	0.091
	Northeast	110,889	11.23	4947	9.32	−0.197	0.023	−0.174	−0.219	−0.270
	East	95,795	9.70	2725	5.13	−0.663	0.052	−0.611	−0.716	−0.919
	Southeast	126,716	12.83	1872	3.53	−1.333	0.108	−1.225	−1.440	−1.864
	South	155,703	15.77	4716	8.88	−0.599	0.083	−0.515	−0.682	−0.828
	Southwest	120,834	12.24	8712	16.41	0.313	−0.051	0.261	0.364	0.423
	West	103,712	10.50	10,998	20.71	0.736	−0.128	0.608	0.863	0.979
	Northwest	127,984	12.96	10,804	20.35	0.484	−0.093	0.390	0.577	0.650
Elevation	3869–4200	30,654	3.15	304	0.58	−1.738	0.028	−1.710	−1.765	−2.442
	4200–4500	143,803	14.77	6284	11.98	−0.220	0.034	−0.186	−0.254	−0.302
	4500–4800	256,027	26.30	15,152	28.89	0.100	−0.038	0.062	0.137	0.136
	4800–5100	266,770	27.40	19,377	36.94	0.319	−0.148	0.170	0.467	0.431
	5100–5400	157,095	16.13	9563	18.23	0.130	−0.027	0.103	0.156	0.176
	5400–5700	80,207	8.24	465	0.89	−2.279	0.082	−2.197	−2.360	−3.216
	5700–6000	37,286	3.83	1262	2.41	−0.486	0.016	−0.470	−0.501	−0.671
	6000–6357	1819	0.19	47	0.09	−0.764	0.001	−0.763	−0.765	−1.060
Lithology	Hard rock	101,483	10.28	21,463	40.42	1.671	−0.439	1.232	2.110	2.110
	Harder rock	176,351	17.86	1726	3.25	−1.749	0.174	−1.575	−1.923	−2.458
	Softer rock	402,465	40.76	17,828	33.57	−0.204	0.122	−0.083	−0.326	−0.280
	Soft rock	239,235	24.23	11,185	21.06	−0.148	0.043	−0.104	−0.191	−0.202
	extremely soft rock	67,761	6.87	901	1.69	3.261	0.056	3.205	−3.317	−4.591
Distance to faults	0–300	142,236	14.41	947	1.78	−2.138	0.146	−1.992	−2.284	−3.014
	300–600	130,353	13.20	1655	3.12	−1.486	0.117	−1.370	−1.603	−2.083
	600–900	116,897	11.84	3214	6.05	−0.698	0.067	−0.631	−0.766	−0.968
	900–1200	99,413	10.07	4004	7.54	−0.303	0.029	−0.274	−0.333	−0.417
	1200–1500	81,447	8.25	3995	7.52	−0.097	0.008	−0.089	−0.105	−0.133
	1500–1800	62,792	6.36	4748	8.94	0.364	−0.030	0.335	0.394	0.492
	1800–2100	49,558	5.02	4494	8.46	0.562	−0.039	0.523	0.601	0.754
	>2100	304,634	30.85	30,045	56.58	0.655	−0.486	0.169	1.141	0.875
Distance to roads	0–300	75,224	7.62	5010	9.43	0.227	−0.021	0.206	0.248	0.308
	300–600	60,303	6.11	4832	9.10	0.427	−0.034	0.393	0.461	0.575
	600–900	54,734	5.54	5715	10.76	0.718	−0.060	0.658	0.778	0.957
	900–1200	47,437	4.80	6196	11.67	0.972	−0.079	0.893	1.051	1.280
	1200–1500	38,515	3.90	4871	9.17	0.935	−0.060	0.875	0.994	1.234
	1500–1800	35,040	3.55	4053	7.63	0.833	−0.046	0.788	0.879	1.105
	1800–2100	33,165	3.36	3413	6.43	0.702	−0.034	0.668	0.736	0.936
	>2100	642,874	65.11	19,012	35.80	−0.623	0.659	0.035	−1.282	−0.863
Distance to river	0–300	200,667	20.33	5901	11.11	−0.629	0.116	−0.513	−0.745	−0.871
	300–600	141,756	14.36	6769	12.75	−0.125	0.020	−0.106	−0.145	−0.172
	600–900	125,416	12.70	8599	16.19	0.258	−0.043	0.215	0.302	0.350
	900–1200	114,948	11.64	8283	15.60	0.312	−0.048	0.264	0.360	0.422
	1200–1500	102,535	10.39	6996	13.17	0.253	−0.033	0.220	0.287	0.343
	1500–1800	84,136	8.52	5642	10.63	0.235	−0.025	0.210	0.259	0.318
	1800–2100	66,091	6.69	3535	6.66	−0.006	0.000	−0.005	−0.006	−0.008
	>2100	151,738	15.37	7377	13.89	−0.106	0.018	−0.088	−0.125	−0.146

. The change in the calculated weight of each factor at each stage is shown in Figure 4 by the IVM model. With respect to slope and aspect, the most sensitive range of slope is 20–30° (0.28) and the west direction has the largest weight at 0.98, indicating that these two classes are most susceptible to landslide occurrence. The most affected lithology is hard rock (2.1). The elevation class from 4800 to 5100 m has the highest weight factor (0.43). Among the lithology units, hard rock has a value of 2.11. When the distance to faults is >2100 m, the probability of disaster occurrence is the largest, with the weight of 0.88. The class between 900 and 1200 m shows the highest susceptibility (1.28) for distance to roads. At 900–1200 m, the factor of being away from the rivers has the maximum weight (0.42). Finally, the risk assessment map is divided into five categories based on the natural discontinuity method: very low, low, moderate, high, very high (Figure 5). The proportion of disaster area of each hazard level is calculated by statistical analysis. From very low to very high, the percentages of disaster area are 0.39%, 3.03%, 7.87%, 34.22%, and 54.49%, respectively (

Table 5. The proportion of landslides in different susceptibility classes of the 9 models.

Model	Very Low	Low	Moderate	High	Very High
IVM	0.39%	3.03%	7.87%	34.22%	54.49%
W+	0.23%	2.11%	7.59%	26.20%	63.86%
W−	0.55%	3.93%	10.36%	33.40%	51.76%
IVM-LR	0.42%	1.91%	7.38%	29.84%	60.47%
W+-LR	0.12%	2.31%	9.38%	25.95%	62.25%
W−-LR	0.10%	4.54%	14.50%	33.30%	47.56%
IVM-MLP	0.16%	3.57%	9.26%	34.30%	52.71%
W+-MLP	0.11%	2.65%	8.41%	25.19%	63.63%
W−-MLP	0.17%	4.28%	12.68%	33.41%	49.46%

).

3.2. Landslide Susceptibility Mapping by the WoE Model

The relationship between the factors calculated by W⁺ with the WoE method and disasters is shown in Table 4. The most sensitive range of slope is 20–30° (0.09). In the aspect, the west direction has the largest weight, 0.61. When it is at the height of 4800–5100 m, it is most prone to disasters (0.17). The most affected lithology is hard rock (1.23). When the distance to faults is 1800–2100 m, the probability of disaster occurrence is the largest, with a weight of 0.52. The 900–1200 m group displays the factor of the distance from the roads with the highest probability of disaster (0.89). At 900–1200 m away from the river, the calculated weight reaches the maximum (0.26) (Figure 6a). The final hazard assessment map is divided into five categories based on the natural discontinuity method: very low, low, medium, high and very high (Figure 7a). The percentages of disaster area showing the five classes from very low to very high are 0.23%, 2.11%, 7.59%, 26.20% and 63.86%, respectively (Table 5).

The relationship between the factors calculated using the WoE method W⁻ and disasters is shown in Table 4. The change in the W⁻ calculation weight of each factor at each stage is shown in Figure 6b. The most sensitive slope class is 20–30° (0.32), the west direction has the largest weight (0.86), and the areas at a height of 4800–5100 m are most prone to disasters (0.47). The most affected lithology is hard rock (2.11). When the fault > 2100 m, the probability of disaster occurrence is the largest, with a weight of 1.14. The highest probability of the distance to roads between 900 and 1200 m is 1.05. At 900–1200 m away from the river, the calculated weight reaches the maximum (0.36). The final hazard assessment map is also divided into five categories based on the natural discontinuity method (Figure 7b). From very low to very high, the percentages of disaster area are 0.55%, 3.93%, 10.36%, 33.40%, and 51.76%, respectively (Table 5).

3.3. Landslide Susceptibility Mapping by IVM-LR and WoE-LR Models

The relationships between the corresponding factors and the disasters are listed in

Table 6. Coefficient of evaluation factors by LR and MLP.

Factors	Slope	Aspect	Elevation	Lithology	Distance to Faults	Distance to Roads	Distance to Rivers
LR	0.480	0.716	0.150	0.614	0.702	0.367	0.633
MLP	0.053	0.135	0.144	0.164	0.217	0.131	0.157

by logistic regression (LR). Through the coefficients of the LR, the weight of factors from IVM and WoE were weighted superposition. Then, the results of IVM-LR and WoE-LR were divided into five levels using the natural breakpoint method: very low, low, moderate, high, very high (Figure 8).

Based on the IVM-LR model (Figure 8a), the LSI values of all units were between −8.93 and −3.83. Statistical analysis revealed that the proportion of landslide areas in each hazard level from very low to very high was 0.42%, 1.91%, 7.38%, 29.84%, and 60.47%, respectively (Table 5).

Based on the W⁺-LR model (Figure 8b), the LSI values of all units were between −4.64 and −2.10. Statistical analysis revealed that the proportion of landslide areas in each hazard level from very low to very high was 0.12%, 2.31%, 9.38%, 25.95%, and 62.25%, respectively (Table 5).

Based on the W⁻-LR model (Figure 8c), the LSI values of all units were between −5.75 and −3.55. Statistical analysis revealed that the proportion of landslide areas in each hazard level from very low to very high was 0.10%, 4.54%, 14.50%, 33.30%, and 47.56%, respectively (Table 5).

3.4. Landslide Susceptibility Mapping by IVM-MLP and WoE-MLP Models

The factor coefficients obtained by the MLP method are presented in Table 6. The LSI values by the MLP model were obtained through analysis. The results were divided into five levels using the natural breakpoint method: very low, low, moderate, high, very high (Figure 9).

Based on the IVM-MLP model (Figure 9a), the LSI values of all units were between −2.14 and 0.98. Statistical analysis revealed that the proportion of landslide areas in each hazard level from very low to very high was 0.16%, 3.57%, 9.26%, 34.30%, 52.71%, respectively (Table 5).

Based on the W⁺-MLP model (Figure 9b), the LSI values of all units were between −1.36 and 0.57. Statistical analysis revealed that the proportion of landslide areas in each hazard level from very low to very high was 0.11%, 2.65%, 8.41%, 25.19%, 63.63%, respectively (Table 5).

Based on the W⁻-MLP model (Figure 9c), the LSI values of all units were between −1.69 and 0.99. Statistical analysis revealed that the proportion of landslide areas in each hazard level from very low to very high was 0.17%, 4.28%, 12.68%, 33.41%, 49.46%, respectively (Table 5).

3.5. ROC Curves

The accuracy of landslide sensitivity analysis models is usually evaluated by the receiver operator curve (ROC) and area under curve (AUC) [66,67]. Evaluation results are shown in Figure 10. Evaluation accuracies of IVM, W⁺, W⁻, IVM-LR, W⁺-LR, W—LR, IVM-MLP, W⁺-MLP, W⁻-MLP are 86.8%, 86.1%, 87.0%, 87.3%, 86.4%, 86.1%, 86.5%, 86.9% and 85.8%, respectively. The AUC values are 0.868, 0.861, 0.870, 0.873, 0.864, 0.861, 0.865, 0.869 and 0.858 respectively. The results indicate which method is more suitable for working areas at high altitudes. The accuracy rate obtained by the five calculation models was greater than 85%, and the accuracy rate of IVM-LR was the highest.

4. Discussion

In the results of the models, the most sensitive slope strata are 20–30°, indicating that shear stress increases with an increasing slope. The west direction has the largest weight due to the general orientation of the geological layers. Of the elevations, 4800–5100 m has the highest weight factor. The most affected lithology is hard rock, as the dominant lithology of the high elevation is hard bedrock. The highest susceptibility of the distance to faults is above 2100 m because the area the faults are stable, 900–1200 m away from the roads and rivers with the highest susceptibility, indicating that disasters have a significant impact on road and river traffic. Therefore, the results show that landslide susceptibility evaluation is important for disaster prevention and mitigation measures of local governments.

By comparing the percentage of disaster area included in each risk level of the model evaluation results, the percentages of disaster areas in high and very high areas are as follows: IVM is 88.71%, W⁺ is 90.06%, W⁻ is 85.16%, IVM-LR is 90.30%, W⁺-LR is 88.20%, W⁻-LR is 80.85%, IVM-MLP is 87.01%, W⁺-MLP is 88.82% and W⁻-MLP is 82.87%. At the same time, with the increase in hazard level, the landslide area ratio shows an upward trend, which indicates that landslide sensitivity classification obtained by the seven calculation models is reasonable, and the mixed model is more reasonable than single models [68].

It can be seen from the success rate curve that the calculation accuracy of the nine models is more than 85%. In contrast to other regional studies, Cui, et al. used the WoE model to determine the landslide susceptibility mapping of Long County. The AUC method revealed accuracies of 79.71% [69]. Zhu, et al. chose Yadong County as the study area for landslide susceptibility with the LR model. The AUC method revealed accuracies of 83.8%. The results of these models indicated that high-altitude landslide susceptibility evaluation using GIS models is feasible [70,71,72,73]. The prediction accuracies of the hybrid model are relatively high. This shows that the calculation results of the model are accurate and reliable, and that the mixed model is more accurate than the single model [74,75,76].

Based on the above analysis, high-altitude landslide susceptibility evaluation is important and feasible. Furthermore, the disaster assessment classification map produced by the hybrid model has the highest rationality (90.3%), which is consistent with the results obtained in other studies. This study also finds more objective distinctions between the different models, which can be used to obtain better model combinations. In most cases, the hybrid model is better than the single model, and the hybrid model can optimize the problems of the single model to a certain extent. For information model and evidence weight, the calculation of the final LSI may deviate from the actual value through simple numerical superposition. The addition of the logical regression model can be used to calculate the coefficient of each factor, which can further highlight the differences in each factor’s role in the final calculation of the LSI, thus obtaining more accurate and reliable evaluation results. In addition, this research also verifies the above-mentioned assumption and achieves the expected results. However, the study area is small and contains less human activity. Therefore, the results may not be suitable for use in landslide susceptibility evaluation in large cities with intense human activity.

5. Conclusions

The investigation of landslides in high-altitude areas presents important challenges. In this study, we employed GIS-based analysis to assess landslide susceptibility in the study area, utilized remote sensing techniques for landslide detection, and applied single and hybrid models such as information volume, evidence weight, and logistic regression for evaluation. This approach can significantly reduce the financial and personnel investments required for landslide investigations. Moreover, by minimizing human factors, the evaluation process ensures objective results.

The susceptibility to landslides was classified into five hierarchic levels: very low, low, medium, high, and very high, using the natural breakpoint method. The rationality and accuracy of the results were validated by examining the area ratio of each landslide level and analyzing the receiver operating characteristic (ROC) curve. The findings demonstrated that the calculated results were highly reasonable and accurate, with the hybrid model outperforming the single models. Specifically, the IVM-LR model achieved a 90.30% ratio in high-risk and very high-risk areas, an AUC value of 0.873, and an accuracy rate of 87.3%, displaying the highest level of rationality and accuracy. This model proved suitable for landslide risk assessment research in high-altitude areas, providing reliable technical and data support for landslide investigations and assessments.

In this research, we further confirmed the superior performance of the hybrid model compared to single models, using both traditional and machine learning models. Overall, the hybrid model based on machine learning turned out to be less effective than expected. In the future, we will continue to incorporate more machine learning models into the existing framework to enhance evaluation accuracy. This exploration will aid in advancing the field of landslide risk assessment.

In addition, the integration of machine learning models into the existing framework should be explored. Incorporating advanced algorithms and techniques in the field of machine learning can potentially enhance the accuracy and predictive capabilities of landslide risk assessment. Furthermore, future research should focus on incorporating additional influential factors, such as climate change data and geological characteristics, to further improve the understanding and prediction of landslide susceptibility in high-altitude areas.

By continuing to refine and expand the methodology, as well as incorporating new data sources and techniques, we can advance the field of landslide risk assessment. This will contribute to more effective disaster prevention and mitigation strategies, ultimately safeguarding lives, property, and infrastructure in high-altitude regions prone to landslides.