Evaluating Landslide Hazard in Western Sichuan: Integrating Rainfall and Geospatial Factors Using a Coupled Information Value–Geographic Logistic Regression Model

Abstract

The western Sichuan region, characterized by unique geological conditions and the pronounced influence of uneven rainfall patterns, is highly vulnerable to frequent geological hazards, particularly landslides. These events pose significant threats to both public safety and regional ecosystem stability. This study focuses on landslide disasters in Dechang County, Sichuan Province, and introduces a framework for assessing landslide susceptibility. The framework incorporates nine critical factors: slope, aspect, topographic relief, distance from faults, slope structure, lithology, proximity to roads, hydrological systems, and vegetation coverage. Using ArcGIS and integrating rainfall as a key factor, we applied an information value–geographic logistic regression coupled model (GWILR) to evaluate landslide susceptibility across the region. The results show landslide susceptibility in Dechang County is classified into four categories: high (14.02%), moderate (54.06%), low (34.98%), and very low (0.94%). Landslides are most concentrated along fault lines and river systems. The model’s AUC value of 0.926 outperforms the traditional information entropy–logistic regression (ILR) model (AUC = 0.867), demonstrating superior predictive accuracy. The GWILR model offers key advantages over traditional methods. Unlike ILR, it assigns region-specific regression coefficients, capturing spatial heterogeneity and nonlinearity more effectively. The inclusion of rainfall as a key factor further enhances model accuracy by reflecting both temporal and spatial variations in landslide occurrence. This approach provides a more localized and precise evaluation of landslide risk, making it highly applicable for regions with complex geological and climatic conditions. This study highlights the GWILR model’s effectiveness in landslide susceptibility assessment and provides a foundation for improving disaster risk management in Dechang County and similar regions.

1. Introduction

In China’s western mountainous regions, distinguished by their distinctive geological settings and diverse topographical features, there is a high incidence of geological disasters. Numerous urban developments and engineering constructions are situated in areas vulnerable to geological disturbances, such as recessed hillside terrains, debris flow deposition zones, canyon terraces, and inclined slopes, which impose certain constraints on societal advancement [1]. Additionally, the initiation and progression of geological disasters in these areas are significantly influenced by human-induced engineering projects. Activities including excavation of roadside slopes and tunnel advancements can induce mountain instabilities, while construction projects may lead to ecological disruptions and degradation, thereby precipitating or intensifying geological disasters. Evaluating the susceptibility and risk associated with geological disasters entails an in-depth analysis of their occurrence and distribution, taking into account the regional geological structural attributes and topographic conditions. This evaluation is an integral aspect of geological disaster forecasting, offering essential technical support for spatial planning for land utilization, strategies for disaster risk reduction, and engineering measures for mitigation and control [2]. Consequently, the development of predictive models for evaluating the susceptibility and risk of geological disasters is imperative and critical. Such models are intended to offer a scientific and rational framework for the planning and advancement of urban areas in mountainous regions, aiming to significantly mitigate or eliminate human and economic losses.

At present, with the swift evolution of remote sensing and geographic information systems (GIS) technology, a substantial body of research has been devoted to the hazard assessment of geological disasters, achieving noteworthy outcomes. The focus of these studies predominantly lies in dissecting the spatial pattern characteristics inherent to geological disaster susceptibility. From the perspective of research methods, geological hazard susceptibility evaluation methods have gradually changed from qualitative analysis to quantitative measurement [3], Commonly employed methodologies encompass a range of methodologies, like the analytical hierarchical process (AHP) [4], deterministic coefficient methods [5], weight of evidence [6], artificial neural networks [7,8], logistic regression models [9], principal component analysis [10], and information value models [11,12]. Machine learning models primarily include BP neural networks [13], support vector machines (SVM) [14], and random forests (RF) [15]. Despite these advancements, challenges persist in constructing robust evaluation index systems due to potential inaccuracies in prediction models. Artificial neural networks, although highly accurate in prediction, entail complex modeling and evaluation processes and rely on extensive, precise baseline data, which are often challenging to procure at a large scale, limiting their broader applicability [16]. Principal component analysis offers a means to consolidate multiple influencing factors into a singular composite index, effectively reducing collinearity among evaluation indices, yet it overlooks the spatial attributes of these indices. The information quantity method can reveal the impact of varying factor levels on landslides but fails to determine the relative significance of each factor in contributing to landslide occurrences [17]. Logistic regression models adeptly capture the contributions of diverse factors to landslides but fall short in reflecting the influence of distinct factor levels on landslide events [18]. Machine learning models demonstrate high evaluation accuracy and have been successfully applied in landslide susceptibility assessments. However, when conducting assessments over large areas, challenges remain, such as the high demand for database samples and computational power as well as the mechanized classification of influencing factors that do not consider their relationship with landslide mechanisms [19]. The geographic-weighted regression (GWR) model delineates research area boundaries by constructing local regression equations at every point within the spatial domain, aiming to analyze and predict the spatial variability of the subject matter and its driving factors at a given scale [20]. By establishing coupled models to synthesize the advantages of each model to improve the accuracy of the model, predecessors have made more attempts, such as information logistic regression (ILR) [21], deterministic coefficient logistic regression (CFR) [22], a GWR–information-based method [23], etc., and have achieved considerable results, but most of them coupled two models. Its precision and accuracy can be further improved.

In landslide susceptibility assessment models, the logistic regression model, while effective for binary response variables, cannot capture the effects of different factor levels on landslide occurrence. Conversely, the geographically weighted regression model, designed for continuous variables in linear regression, is used to examine linear relationships among continuous factors in susceptibility assessments, but it is not suited for handling binary classification problems. The physical significance of combining logistic regression with GWR to form the geographically weighted logistic regression (GWLR) model lies in its ability to incorporate spatial heterogeneity. It captures the variation in regression coefficients across different geographic locations, reflects the distinct impact and intensity of influencing factors in various regions, and accommodates nonlinear relationships. Local regression analysis in GWLR enhances prediction accuracy, while the addition of spatial weighting allows for a more precise reflection of how geographic environments influence landslide occurrence. Furthermore, as landslide prediction is typically a binary problem (occurrence vs. non-occurrence), GWLR is particularly suitable because it predicts landslide probability through a geographically weighted mechanism, factoring in localized spatial variation. The information value method complements GWLR by capturing the influence of factor gradations on landslide occurrence, providing a means to enhance prediction accuracy and model robustness, especially in regions with complex terrain and geological structures.

Due to the unique geographical location of Dechang County, the region experiences abundant annual rainfall, with concentrated rainfall periods and large amounts of short-duration, heavy rainfall and prolonged, multi-day rainfall events. These rainfall characteristics make the area highly susceptible to landslides and other geological disasters. This study focused on landslides in Dechang County, Sichuan Province, and carefully selected nine factors to construct an evaluation indicator system. By utilizing basic geological disaster data and combining the ArcGIS platform with the incorporation of rainfall evaluation factors, the spatial distribution patterns of the indicators and landslide disasters were analyzed. The information value–geographic-weighted logistic regression (GWILR) coupled model was employed to evaluate the landslide susceptibility and the combined susceptibility considering rainfall factors in the region. This method effectively addresses spatial heterogeneity, enables local regression analysis, and handles complex nonlinear relationships, significantly improving predictive accuracy. It provides innovative technical support for disaster prevention, mitigation, and spatial planning in complex, mountainous urban areas.

2. Overview of the Study Area

Dechang County, situated within the Liangshan Autonomous Prefecture in Sichuan Province, spans a geographical expanse defined by longitudes 101°54′ E to 102°29′ E and latitudes 27°05′ N to 27°36′ N (Figure 1). This area, covering 2284 km², comprises around 10 townships, 2 urban districts, and 65 villages, supporting a permanent populace of 225,000 individuals. The county’s climate is characterized by a subtropical highland monsoon regime. Summers are dominated by warm, moist monsoons from the southwest and southeast, providing abundant warmth, whereas winters are shaped by the influence of polar continental air masses, with the upper atmospheric conditions governed by dry, warm southwesterly winds. Distinct wet and dry seasons mark the region’s climate, with the rainy season extending from May to October and the dry season from November to April.

The research area, along with its adjacent regions, is marked by the presence of active faults. Dechang County, characterized by its intricate topography and geological structures, undergoes vigorous neotectonic movements. Such dynamics have led to the severe fragmentation of geological strata due to intense dissection and compression, undermining mountain stability. This has set the stage for a range of geological hazards, including collapses, landslides, and debris flows, thereby creating a geological environment prone to frequent and severe disasters.

The average annual precipitation in the region is 1089.8 mm, but it is unevenly distributed over time. From May to October, the rainfall reaches 1008.3 mm, while from November to April, it is only 68.5 mm, showing a distinct dry–wet seasonality with a significant disparity in precipitation, often concentrated in intense rainfall events. The maximum, minimum, and average monthly rainfall over the past decade (2013–2022) were compiled, as shown in Figure 2. As indicated by the figure, there is a clear increasing and decreasing trend, with the maximum rainfall in July reaching 433.9 mm. Rainfall is predominantly concentrated between June and September. Hydrologically, the region is part of the upper Yalong River system, a tributary of the Jinsha River. The main rivers in the county include the Anning River, which flows from north to south across the entire area, along with its tributaries such as the Cida River and Jincha River, as well as the Yalong River, which forms the western boundary of the county. The primary landform types in the area can be categorized into four major classes: mid-mountain, mid-high mountain, high mountain, and Anning River valley plains. The region has complex stratigraphy and lithology, with exposed Jurassic and Triassic sandstone, mudstone, and shale containing coal-bearing formations. The rock layers are prone to weathering, with loose structures and low mechanical strength. In the slope zones, the surface is covered by Quaternary loose deposits, including colluvium and debris flow deposits, with thickness varying across different locations, generally ranging from 1.5 m to 3.5 m. The bedrock consists of mudstone and shale, which are semi-hard rocks. The mudstone has a high clay mineral content, making it highly absorbent and prone to softening when in contact with groundwater, forming weak structural surfaces with low shear strength, making the area susceptible to landslide disasters. Investigating and comprehending these dynamics is of paramount theoretical significance and practical relevance to enhancing local strategies for disaster prevention and mitigation.

Based on historical disaster records, remote sensing interpretation, and field verification, it was determined that a total of 295 landslide sites developed in the research domain after 2021. Among these, 290 were soil landslides, accounting for 98.3%, and 5 were rock landslides, representing 1.7%. The landslides in the research area were predominantly characterized as small- to medium-sized soil landslides of the traction type. Some typical landslide field photographs are shown in Figure 3.

3. Research Methods

The construction of the model is a crucial step in landslide susceptibility evaluation. The specific application process of the model in this study is illustrated in Figure 4. To ensure the balance of the data samples, the study selected both positive and negative samples for modeling. Non-landslide points were randomly generated 100 m from the landslide locations. Attributes for each evaluation factor were collected from both landslide and non-landslide points. The study utilized 270 collected landslide materials and 270 non-landslide materials, resulting in a total of 540 positive and negative samples.

3.1. The Information Value Model

The onset of geological disasters is determined by the interplay of multiple factors, whose effects may vary across different geological contexts, thereby suggesting the existence of an optimal combination of factors [24]. In geological hazard assessment, the information value model refers to an indicator that quantifies the level of involvement of various factors to the probability of geological hazards by statistically assessing the relationships between historical geological hazard data and influencing factors. Higher information values indicate an increased probability of landslide occurrence. By integrating data on geological hazard points with the spatial distribution of evaluative factors, the information value method provides a practical and straightforward quantitative assessment tool for the vulnerability zoning of geological hazards [25]. When I > 0, there are favorable conditions for a landslide. When I < 0, conditions are not conducive to landslides.

$$I=\ln {\displaystyle \frac{N_{ij}/N}{S_i/S}}$$

(1)

where I represents landslide hazard information; N_ij represents the number of landslide disasters in the classification of impact factors; N denotes the total number of landslide disasters in the region; S_i represents the number of graded grids affecting the factor; and S signifies the aggregate count of grid squares within the designated region.

3.2. Logistic Regression Model

Logistic models are frequently employed as a statistical tool for analyzing binary outcomes. In this model, selected factors serve as independent variables to describe the manifestation of geological disasters, where 0 indicates the absence of a geological disaster, and 1 indicates its occurrence. With multiple independent variables (X₁, X₂, X₃, … X_n), this represents a nonlinear classification statistical method [26].

$$\left\{\begin{array}{l}P(Z=1\left|X\right.)={\displaystyle \frac{1}{1+e^{-Y}}}\\ Y={\beta }_0+{\beta }_1{x}_1+{\beta }_2{x}_2+\cdots +{\beta }_n{x}_n\end{array}\right.$$

(2)

In the context of logistic regression, β₀ serves as the constant of logistic regression, whereas the coefficients β₁ through β_n correspond to the logistic regression parameters associated with each respective variable. The term p denotes the likelihood of a landslide occurrence, with x_n signifying the variable at position n.

3.3. Geographicaly Weighted Regression Model

The geographically weighted regression approach, introduced by Fotheringham, enhances the traditional linear regression framework by incorporating spatial coordinates directly into the regression coefficients [27]. This integration allows the GWR model to adeptly handle spatial variability, making it a robust tool for analyzing data with spatial heterogeneity. The formulation of the GWR model is as follows:

$$y_i={\beta }_0(u_i,v_i)+{\displaystyle \sum _{k=1}^p{\beta }_k}(u_i,v_i)x_{ik}+{\epsilon }_i$$

(3)

where y_i represents dependent variable; x_ik denotes the kth explanatory variable of the ith group; i = 1, 2, … n is the number of samples; (u_i,v_i) represents the coordinates of the ith sample point (for example, latitude and longitude coordinates); β_k(u_i,v_i) denotes the kth regression parameter for the ith sample point of the geographical position function; and ε_i is the random error of the ith sample point.

Typically, the regression coefficient β(u_i,v_i) is derived through the application of the least squares technique:

$$\beta (u_i,v_i)=(X^{T}W{(u_i,v_{i}X)}^{-1}{X}^{T}W(u_i,v_i)Y$$

(4)

In this analysis, W(u_i,v_i) represents a diagonal matrix of order n, which is determined by the selected function for spatial weighting. For the purposes of this investigation, the Gaussian function was utilized as the spatial weighting mechanism.

The Gaussian function method is as follows:

$$w_{ij}=\exp \left[-{(d_{ij}/h)}^2\right]$$

(5)

where h represents the bandwidth parameter, and d_ij denotes the measurement of distance from sample point i to sample point j.

3.4. The Coupled Model of Information Volume, Logistic Regression, and Geographic-Weighted Regression (I+ LR+ GWR)

In geographic-weighted regression and logistic regression model coupling, initially, the geographic spatial information of geological disaster points is incorporated into the logistic regression model. Combined with the previously mentioned Equation (2), this yields Equation (6).

$$\left\{\begin{array}{l}P(Z=1\left|X\right.)={\displaystyle \frac{1}{1+e^{-Y}}}\\ Y={\beta }_{0(u_i,v_i)}+{\beta }_{1(u_i,v_i)}{x}_{i1}+{\beta }_{2(u_i,v_i)}{x}_{i2}+\dots +{\beta }_{n(u_i,v_i)}{x}_{in}+\epsilon \end{array}\right.$$

(6)

After the change of the above formula, Formula (7) can be obtained.

$$\log it(\mathrm{P})=\ln {\displaystyle \frac{p}{1-p}}={\beta }_{0(u_i,v_i)}+{\beta }_{1(u_i,v_i)}{x}_{i1}+{\beta }_{2(u_i,v_i)}{x}_{i2}+\dots +{\beta }_{n(u_i,v_i)}{x}_{in}$$

(7)

Subsequently, by solving for the estimated coefficients of the local regression model at geological disaster point i, Equation (8) can be derived.

$$\widehat{\beta }(u_i,v_i)=(X^{T}W{(u_i,v_{i}X)}^{-1}{X}^{T}W(u_i,v_i)\log it(\mathrm{P})$$

(8)

where (u_i,v_i) represents the geographical coordinates of the i geological hazard site; Y represents the dependent variable of geological hazard point I; β₀ denotes a constant term; β(u_i,v_i) denotes estimates of model parameters; ε is random error (negligible in calculation); W(u_i,v_i) is the spatial weight matrix; and X is the explanatory variable matrix.

Finally, on the basis of the above formula, the X explanatory variable matrix is replaced by the calculated information quantity value of each sample point, and the coupled model of the three models is completed [28] to obtain the integrated model of information-driven and geographic-based logistic regression.

3.5. Probability Density and Landslide Area Ratio

This study employed probability density (PD), landslide area, and landslide area ratio (LAR) as indicators to describe the spatial distribution characteristics of landslides in Dechang County [29]. PD refers to the proportion of landslide occurrences within each range of influencing factors relative to the total number of landslides, reflecting the correlation between landslide frequency and variations in each factor. This parameter can be expressed as follows:

$$PD={\displaystyle \frac{\frac{N_i}{N_t}}{C_i}}\times 100\%$$

(9)

where i represents each interval, N_i denotes the number of landslides within that interval, N_t is the total number of landslides in the study area, and C_i represents the interval class width. The interval width for discrete variables is set to 1.

The landslide area ratio indicates the proportion of the area occupied by landslides within each range of influencing factors. The calculation formula for this parameter is as follows:

$$LAR={\displaystyle \frac{A_{class}}{A_{all}}}\times 100\%$$

(10)

where A_class is the area of landslides in that class, and A_all denotes the aggregate area of landslides within the region under investigation.

4. Susceptibility Evaluation

4.1. Evaluation Unit

The intensity of landslide occurrences is subject to a multitude of factors, demonstrating notable variability and complexity within specific locales. Before undertaking a regional assessment of geological disaster susceptibility, it is essential to segment the entire study area into discrete units. In the context of landslide hazard evaluations, a “unit” represents the minimal surface entity under consideration. The strategy employed for delineating evaluation units, along with the dimensions of these units, significantly influences the accuracy of the assessment outcomes. Presently, scholars both within and outside the country have broadly categorized the methodologies for partitioning evaluation units into three main types: regular grid units, natural slope units, and administrative units, each with its implications for the precision of the evaluation [30].

Considering the simplicity and additive benefits of grid cell segmentation, this study employed grid cells as the assessment units for evaluating landslide hazards. The grid size was determined by leveraging empirical formulas from prior research and considering the specific conditions of the study area, leading to a grid resolution of 18 m × 18 m, encompassing a total of 7,120,654 grid cells.

4.2. Selection and Grading of Evaluation Factors

4.2.1. Selection of Evaluation Factors

The incidence of landslide disasters arises from the combined effects of fundamental geological and environmental conditions along with triggering factors [31]. By collecting detailed survey data at a 1:50,000 scale for the study area and based on relevant research findings, a preliminary selection of terrain and geological factors was made in consideration of the current landslide disaster situation in the region. These factors include topographic features such as aspect, slope shape, terrain relief, and slope gradient; geological factors such as lithology and distance to fault; as well as factors such as slope structure, distance to water systems, distance to roads, and vegetation coverage, resulting in a total of 10 evaluation factors. To minimize data duplication and enhance the precision of the assessment model, this study computed the Pearson product–moment correlation. This coefficient ranges between −1 and 1, with an absolute value |R| closer to 1 indicating a stronger linear relationship between the two variables. The Pearson correlation coefficient categorizes the level of correlation into three intervals: 0 ≤ |R| < 0.3 indicates a low correlation between two variables [32]; 0.3 ≤ |R| < 0.8 signifies a correlation of moderate intensity; and 0.8 ≤ |R| < 1 suggests a strong correlation between the variables. In susceptibility assessment studies, if a moderate to high correlation is identified between two evaluation indicators, one of them ought to be eliminated to mitigate the influence of correlation, with the specific Pearson product–moment correlation presented in

Table 1. The matrix of Pearson correlation coefficients for the assessment factors.

	X1	X2	X3	X4	X5	X6	X7	X8	X9	X10
X1	1
X2	0.773	1
X3	0.014	0.009	1
X4	−0.438	−0.404	0.001	1
X5	0.174	0.148	0.007	−0.194	1
X6	−0.164	−0.152	0.017	0.280	0.030	1
X7	0.419	0.371	−0.013	−0.461	0.269	−0.269	1
X8	0.006	0.009	−0.001	−0.022	0.024	−0.002	0.016	1
X9	0.037	0.002	−0.127	−0.023	0.047	0.001	−0.010	−0.001	1
X10	0.216	0.206	−0.096	−0.236	0.062	−0.166	0.258	0.013	−0.078	1

Note: X₁–X₁₀ indicate slope, slope shape, slope structure, lithology, distance from water system, distance from fault, distance from road, terrain relief, slope orientation, vegetation coverage, respectively.

.

Through the analysis of the Pearson correlation coefficient matrix of each evaluation factor, slope factor and slope shape factor have a high correlation degree. In conjunction with the development distribution law of landslide hazard in Dechang County and field investigation and analysis, the slope shape factor is eliminated, and the evaluation factor system of landslide disaster susceptibility in the investigation area is finally obtained: topographic and geomorphic condition factors (slope, slope orientation, and topographic relief). These factors are classified based on their direct influence on slope stability. Steeper slopes, irregular topography, and unstable slope structures increase the likelihood of landslides, and thus, they are classified at higher levels of susceptibility. Geological structural factors such as proximity to faults and the type of lithology play a crucial role in landslide occurrence. Faults represent weak zones that can trigger landslides, and certain lithologies (e.g., soft rocks or fractured materials) are more prone to failure, leading to higher classification levels, while additional factors must also be considered (slope structure, distance from road, distance from water system, and vegetation coverage). These factors are assessed based on their indirect impact. Proximity to roads and rivers often leads to human-induced disturbances or water-related erosion, increasing landslide risk. Based on the ArcGis analysis function, each evaluation factor was quantified into 18 m × 18 m raster data, as shown in Figure 5.

To ensure the accuracy of factor selection, a multicollinearity analysis was conducted. This analysis is an evaluation method used in linear regression to assess whether there is a strong linear association among predictor variables. Severe multicollinearity can distort model estimates or hinder accurate evaluation, resulting in unstable analytical outcomes. Typically, the variance inflation factor (VIF) and tolerance (TOL) metrics are employed to evaluate multicollinearity among factors. In this study, SPSS 27.0 software was employed to perform multiple linear analyses on the evaluation factors. It is generally accepted that a VIF exceeding 2 and a TOL below 0.5 indicate strong multicollinearity between factors; conversely, values below these thresholds suggest the absence of multicollinearity issues. The results, presented in

Table 2. Evaluation factor multicollinearity diagnosis table.

Evaluation Factor	TOL	VIF
Slope	0.647	1.546
Lithology	0.794	1.259
Vegetation coverage	0.788	1.268
Distance from fault	0.928	1.077
Distance from water system	0.723	1.384
Distance from road	0.883	1.132
Terrain relief	0.684	1.463
Slope orientation	0.870	1.149
Slope structure	0.702	1.425

, show that all factors have TOL values greater than 0.5 and VIF values less than 2, indicating no multicollinearity present. This finding validates the reliability and effectiveness of the selected evaluation factors.

4.2.2. Evaluation Factor Importance Judgement

The weighted importance of the various evaluation factors was assessed using the same model. In the 1980s, J. Moody introduced a neural network model referred to as the radial basis function (RBF) network, which comprises a three-layer forward-propagating neural network. The RBF network is a type of local approximation network that demonstrates significant advantages in both input data and output processes. This model can evaluate the significance of features by adjusting each one individually and observing the conclusive prediction outcomes [33].

Therefore, this study utilized the RBF neural network model to assess the relative significance of each evaluation criterion, as illustrated in Figure 6. In the determination of evaluation factor importance, a threshold of 0.05 was selected to ensure that the chosen factors have sufficient statistical significance in relation to landslide occurrence. This value is commonly used in the field of geological disaster research, where factors with an importance value below this threshold are typically considered to have a minimal impact on the results. Additionally, it helps to avoid over-simplification of the model while maintaining its stability and interpretability. The results indicate that all criteria have importance scores greater than 0.05, suggesting that these factors significantly influence the occurrence of landslides within the study area. Consequently, it was decided to include all nine evaluation indicators in the assessment model.

In the ranking of factor importance, the distance to fault lines is crucial for landslide occurrence. As shown in Figure 7, a distribution relationship between the fault lines and landslide points was established. From the figure, it can be observed that the closer the landslides are to the fault lines, the greater the number of landslides. The landslides are primarily concentrated within distances of <1000 m and 1000–2500 m from the fault lines.

The following is a brief introduction to the nine evaluation factors that influence landslide occurrences, and the distribution map of landslide influencing factors is shown in Figure 8.

4.2.3. Topographic and Geomorphological Factors

Slope: Slope is closely related to landslide hazards. A suitable topographic slope is a prerequisite for landslide hazard development. Among the controlling factors, slope factor plays an important role in controlling its formation mechanism and anti-sliding critical point of landslide [34]. The statistical results are illustrated in Figure 8a, where the probability density first increases and then decreases, reaching its maximum value in the 10–20° interval. This range accounts for 108 landslides, representing 36.61% of the total landslides. The regions affected by landslides in the 20–30° and 30–40° intervals are relatively large, measuring 709,425.07 m² and 677,578.15 m², respectively, with the landslide area ratio showing an increasing trend, particularly within the 20–40° range.

Slope orientation: Slopes with different aspects exhibit significant variations in terms of sunlight duration, intensity, and rainfall. As a result, vegetation cover and soil moisture content on the slopes differ, thereby influencing the stability of the slope [35]. By using the terrain analysis feature of ArcGIS10.8 software, slope direction and the characteristics of the study area can be derived from DEM data. As demonstrated in Figure 8b, PD shows a gradually increasing trend, mainly concentrated in the southwest, west, and northwest directions, with landslide counts of 38, 53, and 67, respectively, along with a high LAR.

Topographic relief: Topographic relief pertains to the disparity between the highest altitude and the lowest altitude in a specific region, which is a macroscopic index to describe the topographic characteristics of a region [36]. Dechang County is located in the eastern margin of Kang-Tibet Plateau, a part of Hengduan Mountain Range. It is a middle mountain and middle mountain river terrace. The topography is very undulating, and the disaster type is mainly landslide. In Figure 8c, the statistical results indicate that PD first increases and then decreases, primarily concentrating in the terrain undulation ranges of 50–100 m and 100–150 m, accounting for 46.1% and 36.27% of the total landslides, respectively, while also exhibiting a high LAR.

4.2.4. Geological Factors

Distance from faults: Landslide disasters frequently occur in regions with active faults, and there is a tight linkage between them. Specifically, in the intersections of regional faulted structures, rocks are often more fractured, and this creates structural conditions that facilitate the genesis and evolution of geological hazards [37]. This leads to enhanced fragmentation and weathering of the rock mass, reducing the shear strength of the rock mass. The rocks in the areas near faults are often more porous and have a higher water content, thereby increasing the possibility of landslide occurrence. As shown in Figure 8d, the statistical results exhibit a decreasing trend in PD, peaking in the interval of distance from fault < 1000 m, with 142 landslides (48.14% of the total) and the highest LAR. This is attributed to the fact that geological materials closer to the fault exhibit greater fragmentation and instability, leading to a higher susceptibility to landslides.

Lithology: Stratum lithology and rock and soil type are important material conditions for forming slopes, which determine the structure of slopes and the type characteristics of geotechnical engineering geology, and provide material basis for the occurrence of landslides [38]. The lithology of the study zone includes gravel soil, clay, soft rock, interbedded soft and hard rock layers, moderately hard rock, and hard rock. In terms of lithology, there are five types of rock formations in the landslide distribution area, as shown in Figure 8e. The most frequent landslides occur in the gravelly soil category (Type 1), with 77 occurrences and the largest scale, resulting in the highest LAR value. Analysis indicates that landslides also concentrate in the clay (Type 2) and weak rock (Type 3) categories.

4.2.5. Human Engineering Activity and Other Influencing Factors

Slope structure: Slope structure will directly affect the stability of landslides. Steeper slopes and rock types prone to collapse increase the risk of landslides and also affect the morphological characteristics of landslides [39]. Different slope structures can lead to different types of landslide, such as top slide, base slide, or integral slide. Figure 8f reveals that slope structures predominantly consist of oblique and horizontal slopes, with the number of landslides in these categories being 90 and 96, respectively, representing 30.51% and 32.54% of the total landslides and again reflecting a high LAR.

Distance from roads: The road demonstrates the influence of human engineering interventions on rocks and soil. In the building process of large-scale initiatives, it is an inevitable process to cut hills and slopes. Therefore, due to the influence of vibration and disturbance, voids can easily occur in the rock and soil, which promote the infiltration of water and change its natural stress state. When the sliding body exceeds its equilibrium state, it is easy to cause landslide disasters [40]. The results in Figure 8g show a generally decreasing trend in PD, with the interval of distance from roads < 200 m reaching its peak, comprising 146 landslides (49.49% of the total). This correlates with the increased frequency of human activities that cause vibrations and disturbances to the terrain as proximity to roads decreases, thereby heightening landslide occurrence.

Distance from water system: Water system changes the surface shape, which is the main cause of landslide disaster. The river erodes the hillside on both sides. Under the periodic influence of hydrodynamic force, slopes are prone to depression formation, leading to the gravitational force of the upper rock layers exceeding their maximum tensile threshold, which in turn triggers geological disasters [41]. Figure 8h presents similar statistical trends, where PD decreases with a peak in the interval of distance from water bodies < 200 m. In this interval, 125 landslides account for 48.14% of the total, yielding the highest LAR. Additionally, landslides are concentrated in the 200–400 m range, aligning with the principle that proximity to water systems increases the likelihood of landslides.

Vegetation coverage: Vegetation contributes to soil and water conservation, inhibits hydraulic erosion and erosive activities, and improves slope stability. Its substantial root systems effectively stabilize slopes, thereby mitigating soil and water loss and enhancing resistance to rainwater infiltration [42]. Figure 8i depicts the statistical results, indicating that landslides predominantly occur in the vegetation coverage range of 25–50%, with 117 occurrences, constituting 39.66% of the total. The LAR in this range also reaches its maximum value of 29.70%.This analysis reveals that landslide distribution is concentrated in the 25–50% interval, with an increased probability of occurrence in the ranges of 10–25% and 25–50%.

4.3. Evaluation Model

4.3.1. Calculation of Evaluation Factor Information

Using the information value model described above, each evaluation factor was categorized into multiple levels. The frequency of landslide disaster data points and total zone within each level was then statistically analyzed. The resulting data were substituted into the specified Formula (1) to obtain the information value for each factor category. Utilizing the reclassification function of ArcGIS 10.8 software, the factor layers were assigned values based on their I-value, which served as the new factor weights for subsequent landslide susceptibility overlay calculations. This process ultimately generated a statistical table of information values for landslide susceptibility evaluation factors, as shown in

Table 3. Landslide susceptibility evaluation factor information quantity statistics table.

Evaluation Factors	Grading	Number of Landslides	Area/(km2)	I
Slope (°)	0–10	32	253.6097	−0.1567
	10–20	108	417.8690	0.7202
	20–30	100	709.0662	0.1502
	30–40	45	677.1483	−0.6688
	40–50	10	220.3518	−1.2891
	>50	0	29.0469	0.0000
Lithology	Macadam soil	77	1180.6029	−0.6794
	Clayey soils	68	241.8216	0.8916
	Weak formation	71	545.2619	0.0179
	Soft and hard interbedded rocks	59	234.2588	0.4880
	Harder rock group	20	95.0788	0.6501
	Hard set	0	10.0680	0.0000
Vegetation coverage (%)	0–10	8	152.3273	−1.0740
	10–25	84	419.8395	0.4379
	25–50	117	685.1618	0.3311
	50–75	73	663.9978	−0.1949
	75–100	13	385.7654	−1.3101
Distance from fault (km)	<1	142	1163.9661	−0.0318
	1–2.5	62	679.0722	−0.3174
	2.5–4	42	273.9935	0.1307
	4–5.5	24	115.0741	0.5057
	>5.5	25	74.9859	0.8875
Distance from water system (m)	<200	125	714.5716	−0.3160
	200–400	99	494.5559	0.4784
	400–600	33	375.2526	−0.4001
	600–800	24	270.0336	−0.3938
	>800	14	452.6782	−1.5570
Distance from road (m)	<200	146	405.6603	0.9910
	200–400	55	235.8652	0.7400
	400–600	26	187.8222	−0.1308
	600–800	17	163.3394	−0.1630
	800–1000	12	144.2403	−0.3263
	>1000	39	1170.1646	−1.3211
Terrain relief (m)	<50	41	267.9315	0.0353
	50–100	136	613.6421	0.6084
	100–150	107	933.7431	−0.1042
	150–200	10	426.3471	−2.1032
	>200	1	65.4282	−2.0207
Slope orientation	North	38	291.9091	−0.0197
	Northeast	23	245.9951	−0.6370
	East	21	266.3889	−0.4803
	Southeast	26	267.1723	−0.0954
	South	29	259.8373	−0.2218
	Southwest	38	332.8799	−0.1212
	West	53	338.8133	0.1415
	Northwest	67	304.0960	0.6624
Slope structure	Nearly horizontal stratified slope	3	13.4696	0.6584
	Consequent slope	49	328.8016	0.2149
	Oblique slope	90	762.0441	−0.0937
	Cross slope	96	775.0647	−0.0383
	Adverse grade	57	427.7117	0.0138

.

4.3.2. Coupled Model of Information and Logistic Regression

The essence of the information-based logistic regression model lies in the coupled model obtained through the nesting of the information value approach and the logistic regression approach. The computation process can be articulated as follows: Logistic regression is employed to determine the probability of the dependent variable (whether a landslide occurs), incorporating the weights provided by the information value model to enhance prediction accuracy. Ultimately, the regression coefficients derived from the logistic regression combined with information values allow for the calculation of each factor’s weight, followed by a weighted summation to generate a landslide susceptibility index for the study area.

The implementation steps involve importing the extracted training sample data of each evaluation factor level (including both landslide and non-landslide sample points) into SPSS 27.0 software. The occurrence of landslide disaster points serves as the dependent variable, where 0 represents no landslide, and 1 represents landslide occurrence. The corresponding I-values for each factor level are treated as predictor variables to conduct binary logistic modeling analysis, with the output presented in

Table 4. Table of logistic regression results based on information quantity.

Evaluation Factor (I)	B	Standard Error	Wald	Degree of Freedom	Significance	Exp(B)
Slope	0.541	0.181	8.907	1	0.003	1.717
Lithology	0.922	0.209	19.551	1	0.001	2.514
Vegetation coverage	3.259	0.357	83.502	1	0.001	2.604
Distance from fault	0.551	0.276	3.992	1	0.046	1.175
Distance from water system	0.372	0.162	5.281	1	0.022	1.450
Distance from road	0.521	0.154	11.398	1	0.001	1.685
Terrain relief	0.341	0.099	10.154	1	0.001	1.369
Slope orientation	2.100	0.351	35.810	1	0.001	8.166
Slope structure	1.320	0.403	10.752	1	0.001	3.743
Constant	−0.563	0.112	25.235	1	0.001	0.569

. Here, B represents the weight of each factor, and sig indicates significance; when sig < 0.05, it indicates that the equation holds statistical significance.

The results based on the information-based logistic regression model were also validated by 20% of the test samples to reflect the differences between the approaches, and the ROC curve findings are presented in Section 4.6.

4.3.3. Geographic Logistic Regression Coupled Model Based on Information Value

The geographic logistic regression coupled model further integrates the spatial weighting concept of geographic-weighted regression based on logistic regression analysis, accounting for the heterogeneity of evaluation factors at different spatial locations. GWR 4.0 software was utilized for the spatial heterogeneity analysis and regression coefficient calculations in this step.

The implementation steps consisted of importing the training sample data (the factor levels and corresponding information values for both landslide and non-landslide points) into GWR 4.0 software. The information values of every assessment criterion level served as independent variables, while the incidence of landslide disasters (1 for yes, 0 for no) served as the dependent variable, facilitating geographic logistic regression analysis. This process yielded results comparing traditional global regression (SLR) with geographic logistic regression. For the global regression model, GWR 4.0 first facilitated a global logistic regression analysis, assuming that regression coefficients across the entire study area are constant and do not vary by spatial location, with results displayed in

Table 5. Global regression model parameters.

Evaluation Factor	Est	SE	T (Est/SE)	Exp (Est)
Constant	−0.124180	0.199873	−0.621296	0.883220
Slope	−0.102606	0.203942	−0.503116	0.902482
Lithology	0.196594	0.233610	0.841549	1.217250
Vegetation coverage	0.261637	0.166720	1.569322	1.299055
Distance from fault	1.106076	1.148344	0.963192	3.022476
Distance from water system	0.583286	0.472713	1.233911	1.791917
Distance from road	0.103974	0.164810	0.630874	1.109572
Terrain relief	0.285189	0.273195	1.043904	1.330014
Slope orientation	0.306144	0.328090	0.933110	1.358178
Slope structure	0.044311	0.571253	0.077569	1.045308

. For the geographic logistic regression model, GWR 4.0 further incorporated geographic weighting methods, allowing the regression coefficients at each spatial location to adaptively adjust, thereby capturing the spatial heterogeneity of the relationship between factors and landslide occurrence, as shown in

Table 6. GWILR model coefficient statistics table.

Evaluation Factor	Mean Value	Standard Deviation	Minimum Value	Maximum Value	Median
Constant	−0.152355	0.304136	−0.878577	−0.022787	−0.024808
Slope	−0.099578	0.304136	−0.342678	−0.056742	−0.057254
Lithology	0.255202	0.248253	0.150276	0.848169	0.151518
Vegetation coverage	0.335992	0.283599	0.215518	1.016141	0.217614
Distance from fault	0.931771	0.211132	0.427082	1.026804	1.023578
Distance from water system	0.391632	1.477982	−3.136100	1.018129	1.015768
Distance from road	0.038113	0.241402	−0.538846	0.140262	0.139979
Terrain relief	0.333138	0.485838	0.129130	1.494882	0.129449
Slope orientation	0.321061	0.498628	−0.868219	0.533894	0.532260
Slope structure	0.308777	1.734120	−0.423775	4.446338	−0.421045

.

The Akaike information criterion corrected (AICc) is a diagnostic metric for assessing the fit of the geographic-weighted model; generally, a smaller AICc value indicates better model fit. The results indicate that the AICc value for the global regression is 421.437813, while the AICc value for the geographic-weighted regression is 410.333639, suggesting that the GWILR model, which considering geographic spatial locations, outperforms the SLR model.

The information-based geographic logistic regression coupled model assigns weights to evaluation factors through the information value model, integrating logistic regression with geographic-weighted regression to create a predictive model capable of capturing spatial heterogeneity for landslide susceptibility. GWR 4.0 software is utilized in this process to achieve dynamic adjustments of spatial weight distribution and regression coefficients, thereby enhancing the local adaptability and accuracy of the model.

4.4. Landslide Disaster Susceptibility Evaluation Results

The evaluation method of the information-based geographically weighted logistic regression coupled model was applied to assess the landslide susceptibility in Dechang County. The findings show that the high-susceptibility areas cover approximately 298.92 km², accounting for 12.99% of the surveyed area, predominantly spread along the left bank of the Yalong River and the banks of the Cida River in the high and middle mountain gorges. These areas are characterized by steep slopes, high soil saturation, and tectonic activity, which make them highly prone to landslides. The medium-susceptibility areas cover about 1188.04 km², making up 51.62% of the surveyed area, primarily located in Heilongtan Town, Rehe Town, Cida Town, and various major river basins. These areas, though less prone to landslides compared to high-susceptibility zones, still face a moderate risk due to factors such as slope steepness, proximity to hydrological systems, and varying rainfall patterns. Localized landslides are more common in these areas, particularly after heavy rainfall events. The low-susceptibility areas span approximately 792.64 km², constituting 34.44% of the surveyed area. These areas are typically characterized by gentler slopes, stable geology, and good drainage conditions, all of which significantly reduce the likelihood of landslides. The very low-susceptibility areas are comparatively minor, comprising 0.95% of the surveyed area. These regions are concentrated in urbanized areas within Dechang County, where human activity has greatly modified the natural landscape. The stable geology, low slope gradients, and well-maintained infrastructure have further reduced the landslide risk. Additionally, these areas are subject to effective land management and drainage measures, contributing to their minimal susceptibility. The landslide susceptibility map is shown in Figure 9, and the evaluation results are presented in

Table 7. Landslide susceptibility evaluation results for Dechang County.

Susceptibility Division	Area (km2)	Area (%)	Landslide (Points)	Landslides (%)	Ratio (R = L/A)
High	298.92	12.99	161	54.57	4.20
Medium	1188.04	51.62	112	37.97	0.74
Low	792.64	34.44	21	7.12	0.21
Very Low	21.87	0.95	1	0.34	0.36

.

4.5. Comprehensive Susceptibility Evaluation of Landslide Hazards

Due to the unique geographic location of Dechang County, it experiences abundant annual rainfall with concentrated rainfall periods and volumes, characterized by short-duration heavy rainfall events and continuous multi-day rainfall. The substantial rainfall is a key factor in triggering geological disasters such as landslides. The relationship between the temporal distribution of landslides and rainfall characteristics is shown in Figure 10. As depicted in the figure, the temporal distribution of rainfall directly influences the frequency of landslide occurrences. Rainfall is mainly concentrated between June and September, with July and August being the peak months for landslide disasters, accounting for 64.7% of the total occurrences.

To illustrate the significant impact of rainfall on landslides, a statistical analysis was conducted to examine the relationship between the distribution and development of landslide disasters and the monthly cumulative average rainfall, as shown in Figure 11. The analysis indicates a clear spatial correlation between the distribution of landslide disasters and the monthly cumulative average rainfall. An isopleth map of the cumulative rainfall during the wet season in the study area was created, with seven rainfall categories: ≤93.2 mm, 93.2–97.8 mm, 97.8–102.4 mm, 102.4–107.1 mm, 107.1–111.7 mm, 111.7–116.3 mm, and >116.3 mm. The number and area of landslides within each rainfall category were then statistically analyzed. Using the information entropy calculation formula, the information value for each category was derived, as shown in

Table 8. Statistical analysis of information value for rainfall factors.

Evaluation Factor	Grading	Landslide (Points)	Area/(km2)	I
Rainfall (mm)	≤93.2	65	693.76	−0.4513
	93.2–97.8	87	505.89	0.1931
	97.8–102.4	34	245.60	0.1816
	102.4–107.1	25	300.97	−0.3961
	107.1–111.7	19	202.25	−0.3598
	111.7–116.3	43	166.64	0.6873
	>116.3	22	191.97	0.3854

.

To further understand the impact of extreme rainfall events on landslide susceptibility, a more detailed quantification of their influence on model accuracy is necessary. The

Table 9. Design rainfall values for key areas.

Rainfall Frequency	10%	5%	2%	1%
Calculated Rainfall (mm)	192.38	236.25	290.25	333
Actual Rainfall (mm)	156	209.25	242.55	-
Design Rainfall (mm)	156	209.25	242.55	273

below presents rainfall values associated with different return periods, including calculated rainfall values, actual rainfall values, and design rainfall values for various rainfall frequencies. These extreme events, defined by rainfall exceeding 100 mm within 24 h, have a clear influence on the likelihood of landslides, especially in high-susceptibility zones characterized by steep slopes and loose materials. The calculated rainfall values reflect the expected amount of rainfall based on historical data and return periods, while the actual rainfall values represent the observed rainfall measurements for each corresponding period. The design rainfall values are used in engineering practices to account for extreme events, reflecting the intensity and frequency of rainfall that could lead to landslides.

After quantifying the rainfall factor, it was superimposed with the landslide susceptibility obtained from the previous evaluation, resulting in the integrated susceptibility map for Dechang County incorporating the rainfall factor, as shown in Figure 12.

The comprehensive susceptibility evaluation results of landslide disasters indicate that the high-susceptibility areas of landslide disasters in Dechang County cover a region of 323.52 km², comprising 14.02% of the total area of the surveyed area. These areas include the regions around Luchang–Ganhai–Daxiangping, Xiaogao Town–Leyue Town–Jinsha Lisu Ethnic Township–Jinchuan Township–Yonglang Town, and Tielu–Maan–Dawan–Dashan Township, mainly distributed along both sides of the gully and the towns at the bottom of the gully, with a total of 182 landslide disasters developed. Given the high risk, disaster prevention measures in these regions should focus on enhancing infrastructure resilience (e.g., reinforcing roads, bridges, and drainage systems) and implementing early warning systems for rainfall and soil saturation. Additionally, evacuation plans and public awareness campaigns are crucial, especially in high-population towns. The medium-susceptibility areas span 1246.92 km² (54.06% of the surveyed area), located in the southern hills, flatlands, and the central low and middle mountain slopes of Dechang. While these regions are less impacted by steep terrain or seismic activity, moderate risks persist, particularly during heavy rainfall. Mitigation strategies should include land-use planning that limits construction in vulnerable zones, enhanced slope monitoring, and the reinforcement of infrastructure like schools and hospitals with landslide-resistant designs. The low-susceptibility areas cover approximately 714.73 km², representing 30.98% of the research zone, primarily distributed in the lush vegetation areas east of Yinchang and Tiejia Village. Although these areas are less prone to landslides, local geological disasters can still occur. To maintain stability, ongoing vegetation management and soil conservation practices are recommended, alongside monitoring urban development to prevent destabilization of slopes. The very low-susceptibility areas, comprising 0.94% of the surveyed area, are mostly urbanized with stable geology and infrastructure. Although the risk is minimal, continued attention to urban drainage systems and infrastructure resilience is necessary. Overall, integrating landslide susceptibility maps with urban and rural planning, alongside robust disaster preparedness and public education efforts, will be essential for reducing landslide risks across Dechang County. The comprehensive susceptibility statistics table of landslide disasters in Dechang County is shown in

Table 10. Landslide susceptibility assessment in Dechang County incorporating rainfall factors.

Susceptibility Division	Area (km2)	Area Ratio (%)	Landslide (Points)	Percentage of Landslide Points (%)
High	323.52	14.02	182	61.69
Medium	1246.92	54.06	91	30.85
Low	714.73	30.98	20	6.78
Very Low	21.85	0.94	2	0.68

.

4.6. Accuracy Test of Evaluation Results

The ROC curve, widely used to assess the accuracy of areas vulnerable to geological disasters, effectively illustrates the relationship between the specificity and sensitivity of the evaluation model, providing a clear, intuitive, and precise depiction of its outstanding detection performance. Therefore, the ROC curve is widely used in the assessment of susceptibility to geological disasters [43]. The AUC value of the ROC curve represents the area beneath the curve, serving as an indicator for evaluating model performance, with values ranging from 0.5 to 1. The closer the value is to 1, the more the curve protrudes towards the upper left corner, indicating better model performance. In this study, 20% of the validation sample data was imported into SPSS 27.0 to evaluate the performance of the two models. The AUC value of the information value–geographically weighted logistic regression coupled model was 0.926, while that of the information value–logistic regression coupled model was 0.867. The model evaluation results are illustrated in the ROC curve shown in Figure 13. Based on the validation results, the GWILR exhibits enhanced performance and produces more reliable assessment outcomes.

To provide a comprehensive overview of the model’s classification performance and assist in evaluating its effectiveness under different conditions, accuracy, recall, and F1 score were further calculated. The F1 score, which is the harmonic mean of precision and recall, offers a more comprehensive evaluation standard, especially in cases of class imbalance, and serves as a composite evaluation metric. In this study, the accuracy, recall, and F1 score for both the ILR model and the GWILR model were calculated, as shown in

Table 11. Precision, recall, and F1 results.

Model	ILR Model	GWILR Model
Precision	0.875	0.946
Recall	0.65	0.81
F1	0.736	0.924

. The results indicate that the GWILR model demonstrates higher predictive accuracy and performance.

Additionally, to demonstrate the superiority of the model, a comparative analysis was conducted with the random forest (RF) model. The random forest (RF) model, a widely used machine learning approach, was selected for comparison due to its strong ability to handle complex data with non-linear relationships. Similar to the GWILR model, RF is capable of managing large datasets and multiple input variables. However, there are key differences in terms of model performance, data requirements, and computational cost. The performance of the GWILR model, as measured by the AUC value (0.926), outperforms the RF model, which achieved an AUC of 0.891. This indicates that the GWILR model provides more reliable landslide susceptibility predictions in Dechang County, particularly in regions with complex topography and varying environmental conditions.

5. Discussion

5.1. Reliability of GWILR Model and Practical Implications

This study employed the information entropy–geographic logistic regression (GWILR) coupled model, which has shown strong performance in landslide susceptibility evaluation. The model’s robustness was validated through statistical methods such as ROC analysis and AUC, demonstrating its superior predictive accuracy compared to other methods like information entropy–logistic regression (I-LR). The GWILR model excels in handling spatial heterogeneity, assigning independent regression coefficients to different geographical locations. This feature is crucial for regions with complex geological and topographic characteristics, such as Dechang County, where the influence of environmental factors varies significantly across space.

Compared to other machine learning models, such as the random forest (RF) model, regarding data requirements, the RF model typically requires large datasets with diverse features and labeled instances, whereas the GWILR model can function effectively with fewer data points by utilizing spatial and environmental characteristics in its regression analysis. Additionally, the RF model is computationally intensive, requiring significant processing power, while the GWILR model is less demanding, making it more suitable for regional studies with limited computational resources. Overall, the GWILR model’s advantages are in accuracy and computational efficiency, particularly in data-limited regions, suggesting that hybrid approaches combining the strengths of both models could further improve landslide susceptibility assessments in diverse geological contexts. These models, particularly deep learning approaches such as convolutional neural networks (CNN), offer advantages in automatically extracting complex patterns and relationships, which could enhance prediction accuracy. However, the transparency and interpretability of the GWILR model remain crucial, especially for decision makers who need to understand the underlying mechanisms behind landslide susceptibility. Future work should consider hybrid models that combine the strengths of traditional statistical methods and machine learning models to further improve both predictive power and interpretability.

In addition to its scientific validity, the practical application of the GWILR model is essential for effective disaster risk management. By integrating rainfall as a key factor, the model provides an accurate assessment of landslide susceptibility in Dechang County. Given the substantial rainfall during peak months (June to September), which significantly triggers landslides, the results can be directly applied to inform risk zoning and disaster mitigation strategies. For example, areas identified as high-risk zones, such as the banks of the Anning River and its tributaries, should be prioritized for geotechnical surveys, slope stabilization, and infrastructure reinforcements.

The susceptibility map is also valuable for infrastructure planning, particularly in terms of mitigating landslide risks in vulnerable regions. For example, in the medium- and high-susceptibility zones, infrastructure such as roads, bridges, and residential buildings can be designed with built-in safeguards like retaining walls, slope stabilization measures, and drainage systems to reduce the likelihood of landslide damage. For example, one application scenario is the town of Cida, identified as a high-susceptibility area, which is in need of a new transportation corridor. With the landslide susceptibility map, engineers can design the route to avoid the steepest and most unstable slopes. Additionally, protective infrastructure such as slope stabilization through geo-textiles or retaining structures can be included in the design, thereby reducing future landslide risks and improving the long-term resilience of the infrastructure.

5.2. Challenges and Future Directions for Landslide Risk Management

Due to the unique geographic location of Dechang County, it experiences abundant annual rainfall with concentrated rainfall periods and volumes, characterized by short-duration heavy rainfall events and continuous multi-day rainfall. This substantial rainfall is a key factor in triggering geological disasters such as landslides. By incorporating the rainfall factor, a comprehensive susceptibility evaluation was conducted. Compared to the susceptibility evaluation using nine factors, the area of low susceptibility decreased, while the areas of moderate and high susceptibility increased. Additionally, the number of landslides in the high-susceptibility areas also increased, as shown in Figure 14. This demonstrates the significant influence of the rainfall factor in the evaluation model. Moreover, the temporal distribution of rainfall directly affects the frequency of landslide occurrences.

Additionally, future climate change is likely to exacerbate rainfall patterns in the region, with an expected increase in both the frequency and intensity of extreme rainfall events. Projections suggest that Dechang County may experience more frequent heavy rainfall episodes, which could lead to a higher occurrence of landslides in the future. Incorporating climate change scenarios into landslide susceptibility models will be crucial for enhancing the accuracy of future predictions and developing adaptive strategies for disaster risk management. As such, the integration of climate models and long-term rainfall data will be important in refining landslide susceptibility assessments and ensuring the resilience of communities in landslide-prone regions.

Regarding different geological conditions, in karst landscapes, where terrain is shaped by the dissolution of soluble rocks, slope stability is influenced by factors like cave systems, underground water flows, and rock dissolution. While the GWILR model’s integration of geological and topographic data can identify areas prone to collapses or sinkholes, its performance may be limited by the lack of detailed subsurface data. Future research could benefit from incorporating specialized geophysical methods to assess these subsurface characteristics more accurately. In permafrost regions, landslides are typically triggered by the thawing of frozen soils, leading to rapid changes in slope stability. To improve the GWILR model’s effectiveness in such areas, additional climate and thermal data, along with temporal freeze–thaw cycle information, would be necessary to capture seasonal variations in landslide susceptibility.

Loess plateaus, with their loose, fine-grained soils, are highly susceptible to erosion and slope failures. The GWILR model can assess susceptibility in these regions by accounting for soil composition and slope gradient, though its accuracy could be enhanced by including soil moisture content data, as loess soils are sensitive to precipitation and moisture fluctuations. While the GWILR model is flexible enough to handle different environmental data, its adaptability to complex geological structures needs careful evaluation. The assumption that relationships between factors remain stable across geological settings can limit its applicability. Future work should aim to incorporate site-specific data, including geological and geotechnical parameters, to improve the model’s performance in regions with more complex geological conditions.

While the GWILR model offers a comprehensive framework for landslide susceptibility assessment, the translation of model results into practical disaster risk reduction measures faces several challenges. Key among these is the need to integrate model outputs with existing local government policies and infrastructure planning. Collaboration with local authorities is critical for translating the susceptibility maps into actionable strategies, such as targeted investments in slope stabilization, road maintenance, and early warning systems.

One of the major practical challenges is the variability in landslide triggering factors, particularly rainfall. To address this, future research could focus on integrating real-time meteorological data into the model, creating a dynamic early warning system for landslides. By monitoring rainfall patterns and using model results to issue timely alerts, this system could significantly reduce the response time and increase the effectiveness of local disaster management.

Moreover, integrating socio-economic factors, such as population density and economic vulnerability, into the model could further enhance its application. This would allow decision makers to prioritize risk reduction efforts not only based on environmental factors but also on social and economic resilience. Future studies could explore a more comprehensive risk assessment framework that combines environmental, social, and economic data to guide resource allocation and long-term planning.

6. Conclusions

(1)

Based on the unique mountainous terrain and uneven rainfall characteristics of Dechang County, a comprehensive landslide susceptibility evaluation framework was developed for the region. This framework includes nine key evaluation factors: slope, aspect, topographic relief, distance from faults, slope structure, lithology, distance from roads, proximity to hydrological systems, and vegetation cover. Rainfall was incorporated as a critical factor, and using the ArcGIS platform alongside the information value–geographic-weighted logistic regression (GWILR) coupled model, both landslide susceptibility and comprehensive susceptibility were assessed. The results indicate that factors such as rainfall, distance from faults, proximity to hydrological systems, and lithology significantly influence landslide susceptibility, with the temporal distribution of rainfall directly determining the frequency of landslide occurrences. Although landslide disasters exhibit a certain spatial uniformity, they predominantly follow a clear “linear” and “belt-like” distribution along river basins, particularly in the Anning, Cida, and Jincha Rivers, as well as the deeply incised valleys of the Yalong River;

(2)

This study demonstrates the effectiveness of the GWILR coupled model in addressing the spatial heterogeneity and nonlinear complexity inherent in landslide susceptibility evaluation. Compared to traditional models, the GWILR model assigns independent regression coefficients to different geographic regions, accurately capturing spatial differences across the study area and improving predictive accuracy. Notably, by incorporating the rainfall factor, the model effectively reflects the rainfall characteristics and their influence on landslide occurrences at various spatial locations, further enhancing its applicability in complex geological and climatic settings;

(3)

Landslide susceptibility evaluation results show that, when compared to the information value–logistic regression (ILR) coupled model, the GWILR model achieves an AUC value of 0.926, outperforming the ILR model (AUC = 0.867), thus demonstrating higher predictive accuracy and reliability. This approach provides strong support for future disaster risk assessments in similarly complex geological regions and offers a scientific basis for practical applications in landslide disaster prevention and land planning by local governments and relevant authorities.