Spatial Coupling and Resilience Differentiation Characteristics of Landscapes in Populated Karstic Areas in Response to Landslide Disaster Risk: An Empirical Study from a Typical Karst Province in China

Abstract

Landslides pose a significant threat to the safety and stability of settlements in karst regions worldwide. The long-standing tight balance state of settlement funding and infrastructure makes it difficult to allocate disaster prevention resources effectively against landslide impacts. There is an urgent need to fully leverage the landscape resources of karst settlements and develop landslide risk prevention strategies that balance economic viability with local landscape adaptability. However, limited research has explored the differential resilience characteristics and patterns of landslide disaster risk and settlement landscapes from a spatial coupling perspective. This study, based on landslide disaster and disaster-adaptive landscape data from a typical karst province in China, employs the frequency ratio-random forest model and weighted variance method to construct landslide disaster risk (LDR) and disaster-adaptive landscape (DAL) base maps. The spatial characteristics of urban, urban–rural transition zones, and rural settlements were analyzed, and the resilience differentiation and driving factors of the LDR–DAL coupling relationship were assessed using bivariate spatial autocorrelation and geographical detector models. The key findings are as follows: (1) Urban and peri-urban settlements exhibit a high degree of spatial congruence in the differentiation of LDR and DAL, whereas rural settlements exhibit distinct divergence; (2) the Moran’s I index for LDR and DAL is 0.0818, indicating that urban and peri-urban settlements predominantly cluster in H-L and L-L types, whereas rural settlements primarily exhibit H-H and L-H patterns; (3) slope, soil organic matter, and profile curvature are key determinants of LDR–DAL coupling, with respective influence strengths of 0.568, 0.555, and 0.384; (4) in karst settlement development, augmenting local vegetation in residual mountain areas and parks can help maintain forest ecosystem stability, effectively mitigating landslide risks and enhancing disaster-adaptive capacity by 6.77%. This study helps alleviate the contradiction between high LDR and weak disaster-adaptive resources in the karst region of Southwest China, providing strategic references for global karst settlements to enhance localized landscape adaptation to landslide disasters.

1. Introduction

Karst regions cover approximately 15% of the Earth’s land area and supply drinking water to nearly 25% of the global population, making them one of the planet’s most vital ecosystems [1]. Between 2004 and 2016, complex topography and fragile geological conditions contributed to 4862 landslide events, resulting in 55,997 fatalities, with Asia experiencing the highest spatial concentration of these disasters [2]. Landslide risks in karst regions are further intensified by extreme weather events and human activities. Research indicates that each unit increase in monthly rainfall anomalies corresponds to a 99.7% rise in landslide damage [3]. Agricultural reclamation, urban expansion, and large-scale deforestation increased landslide susceptibility in the Three Gorges Dam reservoir area by 46% between 2015 and 2019 [4]. The karst region of Southwest China is the largest and most intensely developed ecological fragile area among the three major karst plateaus globally [5]. It also experiences the most widespread and severe karst landslides globally [6]. This region is home to approximately 222 million people, primarily engaged in low-productivity agriculture [7], characterized by low socio-economic development and severe human–environment conflicts. In recent years, China has consistently underscored the need to “fortify the national ecological security barrier.” As one of China’s four major ecologically fragile zones, Southwest China’s karst region is pivotal to the nation’s ecological civilization initiatives. Confronted with frequent landslides, limited disaster-adaptive resources, and mounting pressures from national ecological civilization demands, settlements in Southwest China’s karst region must urgently pursue a self-sustaining strategy for landscape adaptation to landslides. However, policymakers’ underestimation of landslide consequences coupled with inadequate risk monitoring and disaster prevention infrastructure due to socio-economic underdevelopment result in 80% of geological disasters occurring outside designated hazard zones annually [8]. A significant information gap exists between landslide disaster risk (LDR) and settlement landscapes in karst regions. Therefore, strengthening the synergy between LDR and disaster-adaptive landscape (DAL) and achieving their spatial integration are essential for enhancing the resilience of landscapes in populated karstic areas to landslide risks and advancing regional ecological civilization efforts.

Extensive research has been conducted on landslide prevention and control. From a risk management perspective, the primary focus is on systematically identifying, assessing, and controlling landslide risks to minimize their negative impact on exposed elements [9,10,11]. On the other hand, resilient cities emphasize their capacity for “reorganization”, allowing for them to absorb, learn, adapt, and rapidly recover from landslide disruptions [12]. Both approaches prioritize proactive disaster adaptation, implementing preemptive measures to mitigate or even prevent disaster-related losses. Response strategies primarily involve landslide vulnerability assessments using various methods [13,14,15], along with terrain monitoring, slope displacement tracking, and remote sensing [16,17,18,19]. Research has also explored disaster early-warning systems triggered by rainfall and soil physical properties [20,21], engineering-based prevention and control methods [22,23,24,25], and emergency response strategies [26,27]. However, all these measures necessitate substantial capital investment, which remains misaligned with the economic underdevelopment of karst settlements. Conventional engineering-based disaster adaptation strategies often reinforce the antagonistic relationship between humans and the environment. Landslide disasters are inevitable; thus, coexistence remains the only viable approach [28]. Resilient landscapes, as integral spatial components of ecological systems, exhibit self-adaptive and self-restorative capacities. When subjected to external disturbances, such as pollution, geological disasters, floods, hurricanes, and droughts, resilient landscapes sustain the stability of ecosystems and landscape communities. Karst regions exhibit a greater diversity of landscape types. Leveraging landscapes as a tool for disaster adaptation presents novel strategies for disaster prevention and mitigation in karst settlements. However, the existing research on resilient landscapes predominantly examines resilience within urban public spaces. Through the establishment of indicator systems [29] and the development of design strategies [30], urban green systems’ ability to withstand and recover from natural disasters has been enhanced. Few studies have investigated the spatially coupled resilience differentiation patterns of landslide risk and settlement landscapes.

In summary, to enhance the disaster-adaptive resilience of karst settlement landscapes and clarify the spatial coupling relationship between landslide disaster risks and settlement landscapes, this study takes Guizhou Province, located in the center of Southwest China’s karst region, as an example. First, the frequency ratio-random forest model and the weighted variance method are applied to generate LDR and DAL base maps and examine their spatial characteristics across three settlement levels: urban, urban–rural transition zones, and rural. Next, bivariate spatial autocorrelation is employed to assess the spatial differentiation of LDR and DAL. Finally, the geographical detector model is utilized to identify the driving factors governing the LDR–DAL coupling. The findings bridge the information gap between LDR and DAL, providing strategic insights to reconcile the conflict between frequent landslide disasters and limited disaster prevention resources in Southwest China’s karst regions. Furthermore, this study proposes a cost-effective strategy for global karst settlements to adapt local landscapes to landslide disasters.

2. Materials and Methods

2.1. Study Area

Guizhou Province lies at the heart of Southwest China’s karst region, encompassing nine prefecture-level cities and spanning approximately 176,000 square kilometers (Figure 1). It is situated in the plateau slope zone, marking the transition from the Yunnan-Guizhou Plateau to the eastern low hills and mountains, characterized by a complex geological structure. It is among the regions in China with the most widespread karst topography. This geomorphological complexity not only shapes Guizhou’s unique natural landscapes but also contributes to significant environmental challenges, as its intricate folds, fault structures, rugged terrain, exposed rock surfaces, and weak soil–rock cohesion make the region highly susceptible to landslides. In recent years, Guizhou’s rapid economic development has intensified human-induced disruptions to the geological environment. The frequency of human-induced landslides has increased, posing a growing constraint on Guizhou’s urban development.

2.2. Data Sources

Geological disaster data are obtained from the spatial distribution dataset of geological disaster points provided by the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences. The digital elevation model (DEM) data are obtained from the CDEMV3 30 m elevation dataset of the National Geospatial Data Cloud and are used to extract indicators, including slope, aspect, planar curvature, profile curvature, terrain roughness, terrain undulation, river network density, and the topographic wetness index (TWI). NDVI data are calculated from Landsat TM8 imagery. Precipitation and fault distribution data are obtained from the National Earth System Science Data Center. Soil erosion data are derived from the mainland China soil erosion dataset compiled by Yan Jining and colleagues on the Scientific Data Platform [31]. Geomorphological type is obtained from the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences. Stratigraphic lithology data are retrieved from the Geological Cloud of the China Geological Survey. Vegetation data are obtained from the vegetation map of the People’s Republic of China [32]. Road network and point of interest (POI) vector data are extracted from the Gaode Open Platform via web crawling. Soil organic matter data are obtained from the National Tibetan Plateau Science Data Center [33].

2.3. Research Methods

As illustrated in Figure 2, this study is primarily divided into two parts: the construction of LDR and DRL base maps and the analysis of resilience differentiation characteristics and coupling driving factors of LDR and DRL. The LDR base map is generated using the frequency ratio-random forest method, whereas the DAL is derived from LDR results integrated with landscape elements through the weighted variance method. This study first examines the spatial characteristics of LDR and DAL across three settlement levels: urban, urban–rural transition zones, and rural. Next, bivariate spatial autocorrelation is applied to assess their resilience differentiation characteristics. Finally, the geographical detector model is utilized to identify the key driving factors influencing the LDR–DAL coupling.

2.3.1. Evaluation Index of LDR and DAL

Drawing on relevant literature on landslide risk assessment [2,34,35] and the characteristics of landslide development in Southwest China’s karst landscapes [36], this study selects indicators from four key domains—topography and geomorphology, climate and hydrology, soil and vegetation, and human activities—to construct a landslide risk evaluation index system. The selected indicators are categorized as follows: Topography and geomorphology—elevation, slope, aspect, surface undulation, terrain roughness, planar curvature, profile curvature, lithology, distance to faults, and landform type; climate and hydrology—average monthly precipitation, river intensity index, and river network density; soil and vegetation—NDVI and soil erosion; human activities—distance to roads and POI density, comprising a total of 17 indicators.

Regarding landslide resistance, vegetation root systems enhance soil stability, mitigate surface water flow, and reduce erosion from rainwater scouring, thereby lowering the likelihood of landslides [37]. Additionally, various vegetation types exhibit differing soil retention capacities, influencing their effectiveness in landslide resistance [38,39]. Studies indicate that vegetation degradation is a major driver of frequent landslides in the mountainous regions of Southwest China [37]. With respect to landslide adaptation, road networks function as natural evacuation routes, buffers, and refuge zones during landslide events [40]. Different road classifications serve distinct disaster mitigation functions [41], and road quality directly influences a city’s capacity to adapt to landslide hazards. Refuge sites are designated areas where disaster victims seek shelter from sudden natural disasters and accidents, equipped with rescue, resettlement, and medical support services [42]. In human settlements, green spaces, plazas, large public venues, and schools serve not only as recreational, aesthetic, and cultural hubs in normal times but also as critical temporary refuge sites during landslide disasters. Regarding landslide recovery, studies suggest that soil organic matter content not only influences the critical moisture threshold—directly affecting landslide occurrence—but also plays a vital role in soil retention and vegetation restoration post-landslide due to its nutrient composition [43,44]. In summary, this study incorporates various vegetation types and soil organic matter from natural landscapes, while roads, park green spaces, plazas, schools, large public venues, and hospitals are selected from artificial landscapes as DAL elements for constructing the DAL base map (

Table 1. Disaster-adaptive landscape indicators.

Refuge Space	Evacuation Route	Vegetation	Soil Organic Matter Content (%)
Park	Highway	Forest land	7.3–14.0
Plaza	National Road	Orchard	14.0–15.4
Museum	Provincial Road	Shrubland	15.4–17.2
School	County Road	Grassland	17.2–23.3
Hospital	Township Road	—	23.3–28.9
—	Other Roads	—	—

).

2.3.2. Frequency Ratio-Random Forest Model

Machine learning is a widely adopted approach for evaluating landslide risk. However, in machine learning-based landslide risk assessments, non-landslide units are typically selected randomly within the study area. This approach may cause non-landslide units to overlap with potential landslide zones, introducing bias into the model’s evaluation results. Thus, a more rigorous selection of non-landslide units is essential, and the frequency ratio-random forest (FR-RF) model effectively addresses this issue [45]. This method first employs the frequency ratio (FR) approach to generate a preliminary landslide risk map. Next, non-landslide units from low-risk areas, selected to match landslide disaster samples, are randomly chosen and integrated with known landslide samples to form the training and validation dataset for the random forest model. Finally, the trained random forest model produces a high-precision landslide risk map, substantially enhancing the model’s predictive performance.

The frequency ratio (FR) calculation formula is as follows:

$$FR={\displaystyle \frac{F_j/F}{C_j/C}}$$

(1)

where $F_j$ represents the number of landslide disaster grids for a given factor within a classification interval, F denotes the total number of landslide disaster grids in the interval, $C_j$ refers to the number of grids for a given factor within the classification interval, and C is the total number of grids in the study area.

Random forest is a sophisticated ensemble classification algorithm [46]. This algorithm utilizes the bootstrap resampling technique, randomly selecting n samples (constituting two-thirds of the total) from the sample set T to construct a training dataset. Each training sample is used to train a decision tree, ultimately forming an ensemble of n decision trees, collectively referred to as a “forest”. During prediction, the final outcome is obtained by aggregating the results of the n decision trees, using majority voting for classification tasks and averaging for regression tasks. In summary, this study employs the frequency ratio-random forest model to construct the LDR base map.

2.3.3. Weighted Variance

The weighted variance method is a statistical technique for quantifying the contribution of different groups or categories to the overall variance in a dataset. This method assigns weights to each group’s variance based on its sample size, reflecting its relative importance in the dataset.

The computation formulas are as follows:

$$W_i={\displaystyle \frac{\sum _{i=1}^k n_i\times \mathrm{var}\left(y_i\right)}{N}}$$

(2)

$$\mathrm{var}\left(y_i\right)={\displaystyle \frac{\sum _{i=1}^n {\left(y_i-{\widehat{y}}_i\right)}^2}{n-1}}$$

(3)

where k denotes the number of classification intervals, n_i represents the sample size of the i-th classification interval, var($y_i$) refers to the variance of the i-th classification interval, ${\widehat{y}}_i$ is the mean of the dependent variable y in the i-th classification interval, and N is the total sample size. This study generates a landslide stability map by inverting the LDR base map and then constructs the DAL base map using the weighted variance method in combination with DAL elements. Here, $W_i$ represents the capability of the settlement landscape to adapt to landslide disasters. A higher $W_i$ value indicates a greater adaptive capacity to landslide disasters and corresponds to a higher DAL level.

2.3.4. Bivariate Spatial Autocorrelation

Spatial autocorrelation measures the degree of correlation between a geographic phenomenon or attribute within a spatial unit and the same phenomenon or attribute in adjacent units. It is typically assessed using global and local spatial autocorrelation metrics [47]. The global Moran’s I index quantifies the spatial clustering tendency of a given attribute within the study area. Its calculation formula is as follows:

$$I={\displaystyle \frac{n}{\sum _{i=1}^n \sum _{j=1}^n W_{ij}}}\times {\displaystyle \frac{\sum _{i=1}^n \sum _{j=1}^n w_{ij}\left(x_i-\widehat{x}\right)\left(x_j-\widehat{x}\right)}{\sum _{i=1}^n {\left(x_i-\widehat{x}\right)}^2}}$$

(4)

where $x_i$ and $x_j$ denote the landslide disaster or disaster-adaptive landscape values of the i-th and j-th spatial units, respectively; $\widehat{x}$ represents the mean attribute value of all spatial units; and $W_{ij}$ refers to the weight matrix between units i and j. The global Moran’s I value generally falls within the range of −1 to 1. A value below 0 indicates a negative spatial correlation, meaning spatial units exhibit dissimilar attributes and a dispersed distribution. A value near 0 suggests a random distribution, while a value above 0 indicates a positive spatial correlation among spatial unit attributes. The statistical significance of the global Moran’s I index is typically evaluated using the p-value.

The local Moran’s I index identifies localized spatial correlation patterns in regional attribute values and is calculated as follows:

$$I_{kl}^i=Z_k^i{\sum }_{j=1}^n W_{ij}{Z}_l^j$$

(5)

where $Z_k^i={\displaystyle \frac{X_k^i-{\widehat{x}}_k}{{\gamma }_k}}$; $Z_l^i={\displaystyle \frac{X_l^i-{\widehat{x}}_l}{{\gamma }_l}}$; $W_{ij}$ represents the weight matrix between spatial units i and j; $X_k^i$ denotes the disaster-adaptive landscape value of unit i, while $X_l^i$ denotes the landslide frequency value of unit i; ${\widehat{x}}_k$ and $X_l^i$ are the mean values of attributes k and l; and ${\gamma }_k$ and ${\gamma }_l$ are their respective variances.

This study investigates the spatial distribution patterns of LDR and DAL through bivariate spatial autocorrelation analysis. Based on the local Moran’s I index, this study classifies areas into four spatial association types: H-H (High LDR, High DAL), L-L (Low LDR, Low DAL), H-L (High LDR, Low DAL), and L-H (Low LDR, High DAL).

2.3.5. Geographical Detector Model

The geographical detector is a quantitative analytical method designed to detect spatial differentiation and identify its underlying driving factors. With advantages such as the absence of linear assumptions, resistance to collinearity, and broad applicability, it has been widely utilized in investigating the factors and mechanisms driving spatial heterogeneity in geographic phenomena [48]. The method compares the within-group variance of an influencing factor across different partitions with the total variance of the entire study region. The calculation formula is as follows:

$$q=1{\displaystyle \frac{\sum _{h=1}^L N_h{\sigma }_h^2}{N{\sigma }^2}}=1{\displaystyle \frac{SSW}{SST}}$$

(6)

$$SSW={\sum }_{h=1}^L N_h{\sigma }_h^2,SST=N{\sigma }^2$$

(7)

where h (1, …, L) denotes the classification of the dependent variable Y or factor X; $N_h$ and N represent the number of units in category h and the entire region, respectively; and ${\sigma }_h^2$ and ${\sigma }^2$ correspond to the variance of Y within category h and the entire region. SSW represents the sum of within-category variances, whereas SST denotes the total variance of the entire region.

This study applies the geographical detector model to examine the driving factors governing the coupling between LDR and DAL. The q value ranges from 0 to 1, with higher values indicating a stronger influence of the factor on the LDR–DAL coupling.

3. Results

3.1. LDR Map

3.1.1. Evaluation of the Frequency Ratio-Random Forest Model

Model performance is assessed using the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, and recall. The AUC value typically falls between 0.5 and 1, with higher values indicating superior model performance. The ROC curve, generated using test samples, is presented in Figure 3, where the model attains an AUC of 0.949. Accuracy measures the proportion of correctly classified landslide and non-landslide units among all samples, with values ranging from 0 to 1. Recall quantifies the model’s ability to correctly identify landslide units, with higher values signifying improved predictive performance. The frequency ratio-random forest model achieved an accuracy and recall rate of 0.91, indicating good predictive performance and its suitability for landslide risk prediction.

3.1.2. Analysis of Risk Results

The landslide risk assessment results derived from the frequency ratio-random forest model are presented in Figure 4. The results reveal that high-risk landslide zones are predominantly concentrated in the central region of the study area, whereas low-risk zones are mainly distributed along the periphery, with high-risk zones covering 42.94% of the total area. At the settlement scale (Figure 5), high LDR zones are more concentrated in urban and urban–rural transition areas, comprising up to 80% of the total, whereas medium and low LDR zones constitute less than 10%. Rural areas demonstrate lower LDR values than urban and urban–rural transition zones, with medium and low LDR levels comprising up to 40%. Overall, low-risk zones in Guizhou Province are predominantly found in rural areas, while urban and urban–rural transition settlements represent critical regions for landslide disaster prevention and mitigation.

Feature importance analysis from the random forest model (Figure 6) reveals that geomorphology, profile curvature, and slope are the three most influential factors in landslide occurrence, with respective importance values of 0.126, 0.118, and 0.113. A comparison of geomorphology, profile curvature, and slope distributions across different settlement types (Figure 7) shows that high-altitude plains and terraces, characterized by low profile curvature and gentle slopes, are more susceptible to landslides than other regions. This risk arises from limestone, the predominant rock type in karst regions, which is highly prone to dissolution by water, resulting in its gradual decomposition. Organic matter in the topsoil is transported by carbonate-rich water solutions into rivers, underground streams, and other hydrological systems. Consequently, steep high-altitude areas become barren as nutrients are transported to karst peak-cluster plains and terraces. Urban areas, however, are typically situated in relatively flat terrain. Compared to rural areas, urban settlements generally feature gentler slopes and thicker soil layers. During heavy rainfall events, increased soil moisture reduces soil strength, rendering urban areas more vulnerable to landslides. In summary, geomorphology, profile curvature, and slope should be prioritized in the planning and development of engineering activities in karst settlements to enhance landslide disaster prevention and mitigation.

3.2. DAL Results

With landslide stability base map data as the dependent variables and DRL elements as independent variables, the weighted variance method was employed to assess the disaster-adaptive capacity of each landscape element. A spatial overlay analysis was subsequently performed to derive the disaster-adaptive landscape (DAL) levels across the study area (Figure 8). The DAL distribution in Guizhou Province presents a fragmented pattern, marked by large-scale dispersion and small-scale clustering in planar and linear formations. This spatial pattern corresponds to the rugged topography and the dispersed, limited flatlands characteristic of karst regions. High DAL areas and above constitute 64% of the total study area, reflecting a generally strong disaster-adaptive landscape capacity. At the settlement scale (Figure 9), high DAL areas constitute less than 5% in urban and urban–rural transition zones, whereas medium and low DAL areas make up as much as 60%. Conversely, over 80% of rural areas belong to the medium and high DAL categories. Compared to urban and urban–rural transition zones, rural areas contain a higher proportion of high DAL regions and fewer low DAL zones. These findings indicate that urban and urban–rural transition zones still require substantial enhancements in DAL development.

From the perspective of DAL components (Figure 10), given that rural areas constitute a significant portion of Guizhou Province’s land area, their DRL levels generally align with the overall average. In urban and urban–rural transition areas, only refuge spaces exceed the average, whereas evacuation routes, soil organic matter, and vegetation remain below average. This suggests that urban and urban–rural transition areas exhibit greater capacity for emergency evacuation following a landslide, whereas rural areas demonstrate stronger preventive capabilities. To support sustainable development, future urban planning should prioritize the development of disaster-adaptive landscape systems, such as evacuation routes and vegetation, while rural settlements should focus on expanding refuge space availability.

3.3. Spatial Autocorrelation Results

Using the GeoDa spatial analysis tool, a spatial weight matrix was developed to compute the global spatial autocorrelation index Moran’s I for LDR and DAL. At a 99% confidence level, Moran’s I was 0.0818, suggesting a positive spatial correlation between LDR and DAL. Among the four spatial association types (Figure 11), H-H is the predominant category in Guizhou Province, comprising 24.4%, followed by H-L at 18.7%. L-L and L-H types are less prevalent, constituting 9.3% and 7.5%, respectively. Overall, Guizhou Province presents a spatial coexistence of high LDR–high DAL and high LDR–low DAL patterns.

At the settlement scale (Figure 12), urban and urban–rural transition areas are primarily classified as H-L, comprising 31%, followed by L-L at 17%, while H-H and L-H collectively constitute only 12%. The H-H and L-H types make up only 12%. This indicates that urban areas generally have low DAL levels, with a significant mismatch between LDR and DAL. As discussed in the LDR analysis, urban and urban–rural transition areas feature relatively gentle slopes, rendering them more vulnerable to landslide risks. Figure 13b,d illustrate that, in urban and urban–rural transition settlements with low DAL, all DAL elements—except refuge spaces—fall below the study area’s average. Conversely, in areas with high DAL (Figure 13a,c), all DAL elements—except vegetation—surpass the average value. This implies that, in urban and urban–rural transition areas, refuge spaces have already reached saturation, indicating that further expansion will not substantially improve overall disaster-adaptive capacity. Instead, strengthening evacuation routes, soil organic matter, and vegetation components can enhance the overall disaster-adaptive capacity, particularly the soil organic matter system, which shows a significant gap from the average. Even in areas with high disaster-adaptive landscape levels, vegetation factors merely reach the average, highlighting the substantial potential for enhancing vegetation landscape systems in urban and urban–rural transition areas.

In rural areas, H-H is the most prevalent type, comprising 26%, followed by H-L at 18%, while L-L and L-H together constitute 16%. This suggests that LDR and DAL are relatively well integrated in rural areas, though certain regions still require further DAL development. A comparison in Figure 13 reveals substantial fluctuations in evacuation routes and soil organic matter components in rural areas. In high DAL regions, these two factors surpass the average, whereas in low DAL areas, they remain below the average. This implies that these two landscape components exert a strong influence on overall DAL levels in rural areas. Additionally, vegetation levels in low DAL areas remain below average, highlighting its role in shaping overall DAL levels in rural regions. Enhancing vegetation coverage in sparsely vegetated rural areas can significantly improve their disaster-adaptive capacity.

3.4. Analysis of Driving Factors

To investigate the driving factors affecting the coupling of LDR and DAL in Guizhou Province, this study employs the coupling coordination degree of LDR and DAL as the dependent variable. Based on prior analysis, potential driving factors are categorized into three domains: disaster-adaptive landscape elements, human activity influences, and landslide disaster risk. Eight indicators were chosen as independent variables: X1 evacuation routes, X2 vegetation, X3 soil organic matter, X4 population density, X5 land use type, X6 geomorphology, X7 profile curvature, and X8 slope. The geographical detector model was utilized to conduct factor detection and interaction analysis, with the results presented below.

According to the single-factor detection results (Figure 14), landslide risk factors exhibit the strongest explanatory power in the coupling of LDR and DAL, followed by disaster-adaptive landscape factors. Among these, slope (X8) and soil organic matter (X3) exert the strongest driving forces on the coupling of LDR and DAL, with values of 0.568 and 0.555, respectively, identifying them as the primary influencing factors. These are followed by profile curvature (X7), vegetation (X2), and evacuation routes (X1), with corresponding values of 0.384, 0.17, and 0.136. Geomorphology (X6), land use type (X5), and population density (X4) exhibit the weakest influence.

In karst regions, limestone readily decomposes. Due to elevation effects, organic matter from the topsoil is transported by erosion and accumulates in low-lying areas. Consequently, areas with gentle slopes and low-profile curvature tend to feature thick soil layers, high organic matter content, and lush vegetation, contributing to elevated DAL levels. Simultaneously, these areas are prone to rapid soil moisture increases during heavy rainfall, which, under gravitational influence, can trigger landslides, leading to high LDR values. In contrast, steeply sloped areas generally exhibit lower soil content, more exposed rock surfaces, and sparse vegetation, frequently leading to both low DRL and low LDR. Therefore, slope, soil organic matter, profile curvature, and vegetation serve as key drivers of DRL–LDR coupling.

The interaction detection results (Figure 15) reveal that the combination of different factors consistently strengthens the explanatory power of DRL–LDR spatial coupling. This effect encompasses 23 instances of bivariate enhancement and four instances of nonlinear enhancement. Slope (X8) and soil organic matter (X3) exhibit strong interactions with other factors. Among these, the interaction between slope (X8) and profile curvature (X7) demonstrates the highest explanatory power, with a value of 0.7105. The second and third strongest interactions occur between soil organic matter (X3) and profile curvature (X7) (0.6841) and soil organic matter (X3) and vegetation (X2) (0.6813). Notably, the interaction between soil organic matter (X3) and population density (X4) further strengthens the explanatory power for risk and spatial coupling.

Single-factor detection results indicate that slope (X8), soil organic matter (X3), profile curvature (X7), and vegetation (X2) are key drivers of LDR–DAL coupling. Consequently, their interactions further amplify this effect. The interactions between slope (X8) and land use type (X5) and slope (X8) and geomorphology (X6) significantly reinforce the explanatory power of the coupling. This phenomenon arises because urban and residential developments in karst regions are often situated in high-altitude plains or terraces with gentle slopes. These areas typically feature abundant vegetation, numerous refuge spaces, and well-developed evacuation routes, fostering strong LDR–DAL coupling. The interaction between soil organic matter (X3) and population density (X4) further reinforces the explanatory power of the coupling. This occurs because areas with high soil organic matter are typically found in karst peak-cluster depressions, which are preferred locations for karst settlements due to agricultural activities and urban development needs. Consequently, areas with high soil organic matter and population density generally exhibit strong LDR–DAL coupling.

4. Discussion

Significant differences exist in land use and ecological functions among urban, urban–rural transition, and rural areas. Therefore, this study classifies karst settlements into three zones—urban, urban–rural transition, and rural—to examine the spatial coupling between LDR and DAL. However, LDR and DAL analysis reveals that urban and urban–rural transition areas share similar spatial differentiation characteristics, whereas rural settlements exhibit significant differences. Consequently, this study primarily focuses on interpreting the LDR–DAL coupling relationship in urban and urban–rural transition areas versus rural settlements.

Local spatial autocorrelation analysis indicates that karst urban settlements generally exhibit low disaster-adaptive landscape levels, with a mismatch between LDR and DAL. High LDR–low DAL is the predominant clustering type. This mismatch is primarily due to the continuous encroachment of urban development on natural resources and engineering activities that reduce the stability of slope structures. Survey data from the Chinese Academy of Sciences’ Huanjiang Karst Ecosystem Observation and Research Station indicate that, in landslide-prone karst cities, vegetation is predominantly grass or shrub-grass (85%), whereas landslides are less frequent in areas with tall trees (6%). Vegetation root systems reinforce soil shear strength, thereby enhancing slope stability. Therefore, integrating residual mountains and parks with suitable vegetation during development can improve soil structure and organic matter content, thereby enhancing soil shear strength and stability in karst settlements. During settlement development, minimizing disturbances to forest systems is essential. Comprehensive assessments should be conducted in areas with poor geomorphic conditions, followed by targeted protection and remediation measures, or development should prioritize geologically stable regions. Implementing these measures to optimize soil organic matter and vegetation coverage can effectively mitigate landslide disaster risks. As a result, the soil organic matter landscape level can be elevated from 3.4 to 4.5, and vegetation levels can be elevated from 5.6 to 6 (Figure 13b,d), leading to an overall 6.77% improvement in disaster-adaptive landscape capacity.

Although rural settlements exhibit better coupling between LDR and DRL compared to urban settlements, 18% still show a concentration of high LDR and low DRL. Rural settlements can mitigate economic losses caused by landslide disasters by reducing development activities in high LDR areas or by focusing on tree planting-based agricultural activities in these areas. This approach would enhance slope stability and reduce the likelihood of landslides.

This study elucidates the spatial coupling between landslide disaster risk (LDR) and disaster-adaptive landscapes (DAL) in karst regions, providing strategic insights for enhancing the localized landscape in karst settlements to adapt to landslide disasters. However, two limitations remain to be addressed in future research. First, this study exclusively examines landslide disaster risk, whereas cities contend with a broader spectrum of complex hazards. Future research should investigate urban landscape adaptability through a multi-hazard lens. Second, this study analyzed the spatial coupling between LDR and DAL at a macro settlement scale; however, socio-economic development and settlement construction exhibit significant disparities across prefecture-level cities in Guizhou Province. Refining the analysis to the county and township levels would facilitate the development of more targeted and effective strategies for strengthening settlement disaster resilience.

5. Conclusions

Effective utilization of localized landscapes in karst settlements can substantially mitigate landslide disaster risks. The coupling of landslide disaster risk (LDR) and disaster-adaptive landscapes (DAL) is crucial for strengthening both economic resilience and disaster adaptability in settlements. Using Guizhou Province, a representative karst region in China, as a case study, this study employs the frequency ratio-random forest model, bivariate spatial autocorrelation, and the geographical detector model to investigate the spatial coupling and resilience differentiation of LDR and DAL in karst settlements. The key findings are as follows:

(1) Urban and peri-urban settlements exhibit a high degree of spatial congruence in the differentiation of LDR and DAL, whereas rural settlements exhibit distinct divergence.

(2) The Moran’s I index for LDR and DAL in Guizhou’s karst settlements is 0.0818, indicating a weak positive spatial correlation. Urban and urban–rural transition areas predominantly exhibit H-L and L-L clustering, whereas rural settlements primarily demonstrate H-H and L-H patterns.

(3) Slope, soil organic matter, and profile curvature serve as the primary drivers of LDR–DAL coupling in karst settlements, with impact strengths of 0.568, 0.555, and 0.384, respectively. The synergistic effects of multiple factors further amplify the explanatory power of spatial coupling. The interaction between slope and profile curvature demonstrates the highest explanatory power (0.7105), particularly in areas with gentle slopes, high soil organic matter content, and low-profile curvature, where LDR and DAL exhibit strong coupling.

(4) In karst settlement development, increasing native vegetation in residual mountains and parks can enhance forest ecosystem stability, effectively mitigating landslide disaster risks and improving the disaster-adaptive landscape level by 6.77%.