Spatiotemporal Heterogeneity and Driving Mechanisms of Rural Resilience in a Karst River Basin: A Case Study of the Wujiang River Basin, China

Abstract

The unique geo-ecological conditions of karst river basins (KRBs) heighten rural vulnerability to compound disturbances; therefore, enhanced rural resilience (RR) is critical for regional ecological security and sustainable development. In this study, the Wujiang River Basin was chosen as the study area. A comprehensive evaluation index system was first established to assess RR. Key driving factors were identified using the Optimal Parameters-based Geographical Detector (OPGD) mode. The Geographically and Temporally Weighted Regression (GTWR) model was then applied to analyze the spatiotemporal heterogeneity in the driving mechanisms of RR. Our results show that from 2010 to 2022: (1) RR in the study area increased significantly, and disparities among counties decreased notably, indicating a trend toward more balanced regional development. (2) RR displayed strong positive spatial autocorrelation, with spatial clusters evolving dynamically under the influence of policy interventions and environmental constraints. (3) The main drivers of spatial heterogeneity in RR included urban–rural income disparity, road network density, agricultural machinery power, etc. Their driving mechanisms exhibited significant spatiotemporal non-stationarity. The findings inform the development of targeted strategies to enhance regional resilience. Additionally, the methodology and empirical insights can serve as valuable references for RR research and practice in other similar KRBs worldwide.

1. Introduction

In the context of rapid globalization and increasing regional development disparities, rural areas—as critical components of the human-earth system-confront environmental challenges such as natural disasters and ecological degradation, as well as developmental pressures including population outflow, economic restructuring, regional inequality, and changes in social structure [1]. With more than three-quarters of the world’s extreme poor living in rural areas [2], enhancing their sustainability is crucial to achieving poverty reduction targets under the 2030 Agenda for Sustainable Development [3]. Against this backdrop, rural resilience (RR) theory has emerged as a crucial analytical framework for sustainable development, providing a novel paradigm to address the adaptive challenges encountered by rural systems under multiple disturbances.

RR is a specific application of resilience theory within rural studies. It crystallized as a distinct academic concept in the early 21st century, concurrent with the rise in socio-ecological systems theory and the evolutionary perspective on resilience [4]. To address the increasingly complex challenges confronting rural areas worldwide, scholars have progressively adopted this framework. Its conceptual evolution parallels the development of resilience theory: shifting from a focus on rural communities’ resistance to and recovery from single risks [5], toward a comprehensive capacity encompassing economic, social, and environmental dimensions, and now emphasizing resilience as a continuous, dynamic process of learning and evolution [6]. This multidimensional and dynamic understanding of resilience is informed by three interrelated theoretical perspectives. Socio-ecological systems theory provides the foundational analytical framework, viewing rural territories as complex wholes in which social and ecological subsystems are tightly coupled [7,8,9]. Complex adaptive systems theory further explains how self-organization, learning, and adaptive behavioral responses to disturbances generate novel system properties [10,11,12]. Coupled human-natural systems theory employs spatially explicit approaches to analyze the manifestations of human–environment interactions within particular geographic contexts, their spatial feedback mechanisms, and pathways for sustainable governance [13,14], thereby providing resilience research with tools for spatial analysis and policy integration. Building on these perspectives, RR is defined in this study as the comprehensive capacity of a rural system. This capacity emerges from the interactions of economic production, social livelihood, and natural ecological subsystems, enabling the system to withstand shocks, adapt to changes, and—when necessary—transform in response to compound natural, economic, and social disturbances, thereby ensuring regional sustainability [15,16].

Existing research on RR has coalesced around three core themes: evaluation frameworks, spatiotemporal evolution, and the influencing factors. Its development exhibits a trajectory from theoretical construction to empirical deepening. For instance, Quandt developed the Household Livelihood Resilience Approach (HLRA) based on the sustainable livelihoods framework, specifically designed for assessing livelihood resilience at the household scale [17]. Alam constructed a resilience index evaluation system for waterfront communities in Bangladesh, a region frequently afflicted by climate disasters [18]. While such studies offer depth and specificity, their context-dependent nature limits transferability, thereby constraining knowledge accumulation and cross-regional comparative analysis. Conversely, evaluation frameworks that strive for universal applicability frequently rely on generic variables. This approach can fail to adequately capture the unique vulnerability structures and adaptive logics of specific localities, leading to a disconnect between scholarly understanding and practical implementation on the ground. Regarding spatiotemporal evolution, research primarily develops along temporal and spatial dimensions. Temporal studies emphasize long-cycle tracking. Salvia conducted a 70-year diagnostic analysis of resilience evolution in agricultural villages in southern Italy [19]. Arouri employed panel data to assess the dynamic impact of natural disasters on the well-being and resilience of rural households in Vietnam [20]. Spatial research, in contrast, strives to reveal the differentiation patterns and geographical mechanisms underlying resilience distributions [21,22]. Their prevalent limitation is the tendency to treat temporal and spatial dimensions separately. Few studies have employed a unified model to reveal how the influence of driving factors evolves across both space and time. Research on influencing factors has established a broad consensus, emphasizing the integrated role of economic, social, and natural capital, and policy interventions [23,24,25]; this has deepened the understanding of resilience. However, at the methodological level, studies still focus on identifying static factors commonly associated with resilience. Whether through case studies or traditional regression models, there is a tendency to assume that driving factors exert homogeneous and constant effects, overlooking their potential spatiotemporal non-stationarity. Consequently, the resulting policy implications often remain at the level of general principles, making it difficult to translate them into differentiated and targeted strategies suited to complex regions.

Karst landscapes are universally recognized as one of the world’s typical fragile ecosystems. In China, they are characterized by representative geomorphic development, vast spatial extent, and acute human–environment tensions, which endow their rural systems with both developmental opportunities and challenges [26,27,28]. This context constitutes a valuable empirical field for in-depth research into the spatiotemporal complexity of resilience-driving mechanisms. Topographically, these regions are characterized by undulating mountains and crisscrossing canyons, with significant elevational variation. Pronounced karst dissolution has also resulted in complex geomorphic features. Ecologically, these areas are marked by shallow, infertile soils and severe rocky desertification. These characteristics support unique ecological landscapes such as peak clusters and caves, which hold high aesthetic value. Conversely, they impose constraints on rural infrastructure development, agricultural production, and daily livelihoods. Therefore, enhancing RR is essential for regional ecological security and promoting sustainability in KRBs.

To address these gaps, this study selects the Wujiang River Basin in China, a typical KRB, as a case study. We quantitatively assess the RR of the study area at four equal time intervals. Using the Optimal Parameter-based Geographical Detector (OPGD) and the Geographically and Temporally Weighted Regression (GTWR) model, we identify the key drivers of RR and investigate the spatiotemporal heterogeneity of their influencing mechanisms. This study aims to achieve the following objectives: (1) To reveal the spatiotemporal evolution and spatial correlation of RR in the study area from 2010 to 2022. (2) To identify the key drivers influencing RR and analyze the spatiotemporal heterogeneity of their driving mechanisms. (3) To provide a scientific basis for policymakers to develop differentiated strategies for enhancing regional RR and promoting sustainable development. The methodology and empirical insights presented here can also serve as a reference for RR research and practice in other similar KRBs worldwide.

2. Materials and Methods

2.1. Study Area

The Guizhou section of the Wujiang River Basin in China (hereinafter referred to as the Wujiang River Basin) is situated in the eastern Yunnan-Guizhou Plateau (26°15′ N–29°21′ N, 104°31′ E–108°07′ E). The topography is predominantly mountainous and hilly, with the terrain higher in the southwest and lower in the northeast. The basin boundary was delineated through literature analysis and spatial overlay techniques, which initially encompassed 44 county-level divisions. To maintain consistency in administrative level and data comparability, 13 municipal districts were excluded, resulting in 31 counties (or districts) as the final study units (Figure 1).

2.2. Construction of Index System

2.2.1. Construction of an Evaluation Index System for RR

Based on the core functions of rural socio-ecological systems [29,30,31], this study constructed an RR evaluation index system structured around three dimensions: agricultural production, residential living, and ecological environment. Informed by the regional context and prior research [32,33,34,35,36], 15 indicators were selected (

Table 1. Evaluation index system for RR.

Objective Layer	Guideline Layer	Indicator Layer	Attributes	Weights
Rural resilience	Agricultural production	Value added of the primary industry	+	0.097
		Per capita resource area in rural areas	+	0.077
		Total agricultural machinery power per unit area	+	0.065
		Grain production per unit area	+	0.031
		Rural human resources	+	0.044
	Residential living	Per capita disposable income of rural permanent residents	+	0.103
		Number of primary and secondary school students per 10,000 people	+	0.025
		Number of beds in medical and health institutions per 10,000 people	+	0.049
		Density of the road network	+	0.167
		Per capita savings deposit balance of residents	+	0.107
	Ecological environment	Carbon dioxide emissions	−	0.019
		Proportion of afforestation area	+	0.072
		Proportion of sloping land area	−	0.052
		Annual average concentration of PM2.5	−	0.069
		Fractional vegetation coverage	+	0.022

Note: The “+” and “−” represent positive indicators (whose increase contributes positively to the overall objective) and negative indicators (whose increase has a negative effect), respectively.

).

The agricultural production dimension was designed to measure the adaptive and transformative potential of the economic subsystem. In the KRBs where arable land resources are constrained, agricultural production efficiency and stability constitute the foundation for withstanding economic and natural risks [37,38]. This study refines the common indicator of “total agricultural machinery power” to “total agricultural machinery power per unit area” to more precisely capture the adaptation of agricultural technology to the region’s fragmented terrain.

The residential livelihood dimension was designed to measure the capacity of the social subsystem to absorb shocks and adapt to change. In KRBs, where rugged terrain impedes public services access, residents’ livelihood security and adequate public services underpin social stability [39,40]. The indicator “density of the road network” was employed to quantify the effectiveness of transportation infrastructure in mitigating topographic barriers and ensuring spatial accessibility.

The ecological environment dimension was designed to diagnose the ecological subsystem’s buffering, regulatory, and self-restoration capacity. In KRBs, where rocky desertification is prominent and the ecosystem is fragile, vegetation cover and soil-water conservation serve as the critical foundation for resisting risks and maintaining stability [41,42]. In response to the core demands of karst ecological restoration, the indicator “proportion of afforestation area” is introduced to quantify active restoration and transformative capacity.

Following the construction of the index system, the indicator weights were determined using the entropy weight method, an objective weighting technique; the specifics of this method are elaborated in Section 2.4.1. Concurrently, to validate the robustness of the weighting system, a cross-verification was performed using the coefficient of variation method. The results demonstrated near-perfect consistency between the weight rankings from the two methods (Spearman’s ρ = 0.989, p < 0.001), which confirms the objectivity and robustness of the weighting system employed in this study. The resulting weights are shown in Table 1.

2.2.2. Preliminary Selection of the Driving Factors for RR

Guided by socio-ecological systems theory and developmental realities of the study area, this study, with reference to the core dimensions of the IPCC [43] and EU Rural Resilience Assessment [44] frameworks, identified 16 driving factors across four dimensions: economic development and financial support, social livelihoods and infrastructure, agricultural production conditions and efficiency, and ecological environment and sustainable development (

Table 2. Selection and calculation of RR driving factors.

Dimensions	Driving Factor	Calculation Formula
Economic Development and financial support	X1 Urban-rural income disparity	Per capita disposable income of urban residents /per capita disposable income of rural residents
	X2 Fiscal self-sufficiency rate	General budget revenue/ general budget expenditure
	X3 Development level of the primary industry	GDP of the primary industry/total GDP
	X4 Economic contribution of the tertiary industry	GDP of the tertiary industry/total GDP
	X5 Loans balance of financial institutions	—
Social livelihoods and infrastructure	X6 Population density	Total population of the region/ total area of the region
	X7 Number of fixed telephone users	—
	X8 Number of beds in social welfare and support units	—
	X9 Density of the road network	Total kilometers of roads/area
Agricultural production conditions and efficiency	X10 Per capita arable land area in rural areas	Cultivated land area/total rural population
	X11 Total agricultural machinery power per unit area	Total power of agricultural machinery/ area sown in crops
	X12 Multiple-crop index	Total sown area of crops/area of arable land
	X13 Cropping structure	Area sown in food crops/total area sown in crops
Ecological Environment and sustainable development	X14 Fractional vegetation coverage	(NDVI − NDVIsoil)/(NDVIveg − NDVIsoil)
	X15 Proportion of afforestation area	Current year’s afforestation area/ total area of the region
	X16 Annual average concentration of PM2.5	—

Note: “X₁–X₁₆” denote the 16 evaluation indicators comprising the resilience index system. “—” is used to denote indicators whose values are directly available from published statistics and do not require a computational formula. NDVI is the normalized difference vegetation index. NDVI_soil represents the NDVI of bare soil or areas without vegetation cover, defined as the NDVI value when the cumulative percentage reaches 5%. NDVI_veg represents the NDVI of fully vegetated areas, defined as the NDVI value when the cumulative percentage reaches 95%

).

The economic development and financial support dimension addressed structural vulnerabilities such as a mono-industrial structure and limited risk-bearing capacity. Indicators, including the “fiscal self-sufficiency rate” and “loans balance of financial institutions” are selected to characterize how local fiscal stability and allocation underpin economic resilience [45]. The social livelihoods and infrastructure dimension examined critical challenges, including underdeveloped infrastructure, inequitable access to public services, and labor shortages. Indicators such as the “number of beds in social welfare and support units” and “number of fixed telephone users” are employed to capture the level of social welfare provision and the connectivity of social networks [45,46]. The agricultural production conditions and efficiency dimension focused on the fragmentation of arable land and low production efficiency. Indicators, including the “multiple-crop index”, are selected to quantify resource utilization efficiency [47,48]. The ecological environment and sustainable development dimension targeted the contradiction between ecological degradation and developmental imbalance. Indicators such as “fractional vegetation coverage” are selected to represent the baseline condition of the ecosystem and its capacity for soil and water conservation [49]. These four dimensions align with the practical challenges and sustainable imperatives for enhancing basin-wide resilience, thereby providing a logical and actionable framework.

2.3. Data Sources and Preprocessing

The data used in this study were obtained from the following sources:

(1)

Statistical data (2011, 2015, 2019, 2023) were collected from the Guizhou Statistical Yearbook, the China County Statistical Yearbook, municipal and prefectural statistical yearbooks, bulletins on national economic and social development statistics, and the China Forestry and Grassland Statistical Yearbook. Missing data were addressed using linear interpolation or imputation with the mean value from adjacent years.

(2)

Spatial and thematic data for the years 2010, 2014, 2018, and 2022: Normalized Difference Vegetation Index (NDVI) data and Land Use and Land Cover Change (LUCC) data were obtained from the Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences (https://www.resdc.cn/, accessed on 15 January 2025), with a spatial resolution of 30 m. Digital Elevation Model (DEM) data, with a spatial resolution of 12.5 m, were sourced from the Geospatial Data Cloud platform (https://www.gscloud.cn/, accessed on 15 January 2025). Hydrological data were derived from the Wujiang River Basin Water Environment Protection Plan (2015–2020). Road network data were acquired from the OpenStreetMap (OSM) website (https://www.openstreetmap.org/, accessed on 15 January 2025), with a spatial resolution of 1 km. CO₂ emission data at a 1 km × 1 km resolution were obtained from the Center for Global Environmental Research (CGER) (https://cger.nies.go.jp/en/, accessed on 16 January 2025). PM2.5 concentration data were provided by the Center for International Earth Science Information Network (CIESIN) at Columbia University, with a spatial resolution of 1 km. (https://ciesin.climate.columbia.edu/, accessed on 16 January 2025).

To ensure effective integration of multi-source spatial data, a standardized preprocessing workflow was applied. First, all raster datasets were georeferenced to a unified projected coordinate system, using the 30 m resolution LUCC data as the spatial reference. Resampling was performed: bilinear interpolation for continuous data and nearest-neighbor interpolation for discrete data. All data layers were clipped to the watershed boundary vector of the study area. This procedure yielded a spatially aligned and scale-consistent foundational dataset.

2.4. Methodology

Figure 2 illustrates the research framework, which is applied to the Wujiang River Basin to analyze the spatiotemporal heterogeneity of RR and its driving mechanisms. The framework comprises the following four main steps:

2.4.1. Entropy Weight Method

The entropy weight method calculates weights based on the information entropy derived from each indicator’s variation across all samples. It mitigates the subjectivity inherent in expert-based approaches and is particularly appropriate for handling multidimensional data.

As indicators vary in dimension and magnitude, data standardization is necessary to eliminate these differences. This study applied distinct standardization methods for positive and negative indicators.

Standardization: Transforming original indicators into dimensionless, comparable values.

For positive indicators:

$$X_{ij}^{\prime }={\displaystyle \frac{X_{ij}-min{X}_j}{max{X}_j-min{X}_j}}$$

(1)

For negative indicators:

$$X_{ij}^{\prime }={\displaystyle \frac{max{X}_j-X_{ij}}{max{X}_j-min{X}_j}}$$

(2)

where $X_{ij}$ denotes the original value of the $j$-th indicator in the $i$-th sample ($i$ = 1, 2, …, m; $j$ = 1, 2, …, n), $X_{ij }^{\prime }$ is the standardized value, and $max{X}_{ij}$ and $min{X}_{ij}$ represent the maximum and minimum values of the $j$-th indicator, respectively.

Probability Calculation: Computing the proportion of each county’s value for a given indicator.

The proportion of the $i$-th sample under the $j$-th indicator is given by:

$$P_{ij}={\displaystyle \frac{{X^{\prime }}_{ij}}{{\sum }_{i=1}^n{X^{\prime }}_{ij}}}$$

(3)

Entropy and Utility Value Calculation: Determining the entropy ($e_j$) and information utility value ($d_j$) for each indicator.

The entropy value $e_j$ for the $j$-th indicator is calculated as:

$$e_j=-{\displaystyle \frac{1}{\ln \left(m\right)}}\sum _{i=1}^m{P}_{ij}·\ln \left(P_{ij}\right)$$

(4)

The information utility value $d_j$ is defined as:

$$d_j=1-e_j$$

(5)

Weight Determination: Deriving the objective weight ($w_j$) for each indicator by normalizing its utility value.

The weight $w_j$ of the $j$-th indicator is then obtained by:

$$w_j={\displaystyle \frac{d_j}{{\sum }_{i=1}^m{d}_j}}$$

(6)

where $m$ is the total number of samples and $n$ is the total number of indicators.

Comprehensive Scoring: the RR score $S_i$ for each sample $i$ is computed using the weighted sum method.

$$S_i=\sum _{i=1}^n{w}_j·X_{ij}^{\prime }$$

(7)

2.4.2. Spatial Autocorrelation Analysis

Spatial autocorrelation analysis is a widely used method for examining spatial dependence and heterogeneity among geographic units. The spatial continuity of KRBs implies that RR is likely spatially dependent. We employed this method to verify its spatial dependence and delineate its specific patterns, such as clusters of high or low resilience.

In this study, Global Moran’s I was employed to assess the overall spatial autocorrelation of RR across counties. A positive value indicates clustering of similar values, a negative value suggests dispersion, and a value near zero implies a random spatial pattern. The formula is expressed as:

$$I={\displaystyle \frac{n·\sum _{i=1}^n\sum _{j=1}^n{W}_{ij}·\left(x_i-\overline{x}\right)\left(x_j-\overline{x}\right)}{\sum _{i=1}^n\sum _{j=1}^n{W}_{ij}·\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^2}}$$

(8)

To further explore local spatial patterns, cluster and outlier analysis (Anselin Local Moran’s I) was applied [50]. This method identifies spatial clusters of high and low values (i.e., high-high and low-low clusters where $I_i$ > 0) and spatial outliers (i.e., high-low and low-high outliers where $I_i$ < 0. The local Moran’s I statistic for each spatial unit is calculated as:

$$I_i={\displaystyle \frac{n·\left(x_i-\overline{x}\right)}{\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^2}}·\sum _{i=1}^n{W}_{ij}\left(x_j-\overline{x}\right)$$

(9)

where $n$ is the total number of spatial units, $x_i$and $x_{j }$ are the RR values of units $i$ and $j$, respectively, $\overline{x }$ is the mean RR value across all units, and $W_{ij}$ is the spatial weight between units $i$ and $j$.

2.4.3. Optimal Parameters-Based Geographical Detector (OPGD)

As an advanced statistical tool, the OPGD improves upon the conventional Geographical Detector (GD) by automatically optimizing key parameters—including the spatial discretization method, number of breakpoints, and spatial scale—to enhance both accuracy and efficiency [51,52]. This study employs both factor detection and interaction detection to quantify the influence of individual drivers and assess their interactive effects.

The factor detection function is defined as:

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

(10)

where $q\in [0,1]$ measures the explanatory power of an independent variable on the spatial variation in the dependent variable. A higher $q$-value indicates a more substantial influence. Here, $h$ = 1, …, $L$ denotes the strata of the variable or driving factor, $N_{h }$ and $N$ are the number of units in the stratum $h$ and the entire region, respectively, ${\sigma }_h^2$and ${\sigma }^2$ represent the variance within stratum $h$ and the global variance.

Interaction detection evaluates whether the combined effect of two driving factors strengthens or weakens their individual impacts on the spatial pattern. The interaction types can be classified into five categories [53], as summarized in

Table 3. Types of interaction between two factors and their specific effects.

Basis for Judgment	Types of Interaction
q(X1 ∩ X2) < Min(q(X1), q(X2))	Nonlinear attenuation
Min(q(X1)), q(X2)) < q(X1 ∩ X2) < Max(q(X1), q(X2))	Single–factor nonlinear attenuation
q(X1 ∩ X2) > Max(q(X1), q(X2))	Dual–factor enhancement
q(X1 ∩ X2) = q(X1) + q(X2)	Independent enhancement
q(X1 ∩ X2) > q(X1) + q(X2)	Nonlinear enhancement

.

2.4.4. Geographically and Temporally Weighted Regression (GTWR)

To quantify mechanisms of spatiotemporal heterogeneity through which the identified drivers influence RR, we employed the GTWR model. While the Geographically Weighted Regression (GWR) model captures spatial non-stationarity by constructing spatially dependent local models, it lacks a temporal dimension. The GTWR model overcomes this limitation by integrating both spatial and temporal coordinates into its weighting matrix, thereby addressing the non-stationarity across space and time [54].

The calculation formula is as follows:

$$Y_i={\beta }_0\left(u_i,v_i,t_i\right)+\sum _{k=1}^p{\beta }_k\left(u_i,v_i,t_i\right)X_{ik}+{\epsilon }_i$$

(11)

where $\left(u_i,v_i,t_i\right)$ denotes the spatiotemporal coordinates of the $i$-th observation, with $u_i$, $v_i$ and $t_i$ representing longitude, latitude, and time, respectively; $Y_i$ is the dependent variable; $X_{ik}$ is the $k$-th independent variable for observation $i$; $p$ is the total number of independent variables; ${\beta }_0\left(u_i,v_i,t_i\right)$ is the local intercept; ${\beta }_0\left(u_i,v_i,t_i\right)$ represents the local coefficient of the $k$-th independent variable; and ${\epsilon }_i$ is the random error term.

3. Results

3.1. Spatiotemporal Characteristics of the Evolution of RR

3.1.1. Temporal Characteristics of the Evolution of RR

The RR of 31 county-level units within the Wujiang River Basin was evaluated for the years 2010, 2014, 2018, and 2022. As shown in Figure 3, the mean RR value increased significantly from 0.206 in 2010 to 0.492 in 2022, with an average annual growth rate of 11.56%. This indicates that over this 12-year period, the overall level of RR significantly improved and maintained a positive development trend. The medians were consistently lower than the mean values throughout the study period, suggesting that although a few high-performing regions elevated the average, RR levels in most counties remained relatively low. Additionally, the coefficient of variation decreased significantly from 0.153 to 0.066, indicating a notable convergence in RR levels across the basin and a narrowing of development discrepancy among counties.

3.1.2. Spatial Characteristics of the Evolution of RR

Using the natural breaks method, RR scores were classified into five distinct levels: low (≤0.221), lower (0.222–0.308), medium (0.309–0.412), higher (0.413–0.486), and high (≥0.487).

Figure 4 visualizes the spatial distribution of RR from 2010 to 2022. A significant increase in the overall RR was observed throughout the study area during this period. Specifically, the extent of low-resilience areas contracted significantly, while the high-resilience regions expanded consistently. Acting as core growth poles, the higher-resilience areas in the northeast and southeast exhibited continuous expansion, evolving from discrete nodal patterns into more contiguous regions. This evolution signifies a positive spillover effect, enhancing RR in adjacent counties. Concurrently, the western and central regions—predominantly characterized by low or lower resilience levels prior to 2014—demonstrated a gradual convergence toward the core areas in terms of both resilience levels and spatial proximity, reflecting a trend toward more balanced regional development.

3.2. Spatial Clustering Characteristics of RR

The results of the global spatial autocorrelation analysis (

Table 4. Global Moran’s I index of RR in the Wujiang River Basin from 2010 to 2022.

	2010	2014	2018	2022
Moran’s I	0.288 **	0.423 **	0.310 **	0.316 **
z-score	2.615	3.907	2.996	3.074
p-value	0.009	0.000	0.003	0.002

Note: ** indicates significant at the 1% level

) show that the Global Moran’s I for RR in the study area from 2010 to 2022 was consistently positive, with all p-values less than 0.01. This indicates a significant positive spatial autocorrelation in RR at the 99% confidence level, reflecting a clear pattern of spatial clustering where neighboring counties influence each other and display similar resilience levels.

A local spatial autocorrelation analysis was further conducted to examine the local clustering characteristics, resulting in a LISA (Local Indicators of Spatial Association) cluster map (Figure 5). The spatial agglomeration patterns of RR from 2010 to 2022 were predominantly characterized by “high-high” and “low-low” clusters.

Between 2010 and 2014, the northeastern region, benefiting from relatively flat terrain and convenient irrigation, exhibited higher resilience levels, forming a “high-high” cluster. In contrast, the central region, constrained by limited arable land resources, was marked by a significant “low-low” cluster. From 2014 to 2018, enhanced efforts in controlling rocky desertification and promoting specialized local industries in the southeastern region led to a marked improvement in RR, resulting in a new “high-high” cluster. Concurrently, increased investment in infrastructure and ecological restoration facilitated a gradual transition in the central region away from the extensive “low-low” clustering pattern. Between 2018 and 2022, RR levels increased significantly across the basin, accompanied by a trend toward more balanced regional development. Spatial outliers (“high-low” and “low-high”) disappeared entirely. The “high-high” clusters became concentrated in the southeastern region, while the “low-low” clusters contracted substantially, persisting only in limited areas of the central region.

3.3. Identification of Key Drivers and Spatiotemporal Heterogeneity of Their Mechanisms

3.3.1. Identification and Interaction Effects of Key Drivers

To identify the optimal spatial stratification, the OPGD model was employed, utilizing four discretization methods (equal interval, geometric interval, quantile classification, and natural breakpoint) across classification tiers ranging from 3 to 9. The classification scheme and method that yielded the maximum q-value were selected as the final configuration (Figure 6).

Using the OPGD model, eight drivers were identified as having a significant influence on RR (

Table 5. Results of factor detection for drivers of RR based on the OPGD model.

	X1	X2	X3	X4	X5	X6	X7	X8
q-value	0.200 **	0.072	0.051	0.084	0.255 **	0.121	0.049	0.162 *
p-value	0.003	0.421	0.471	0.295	0.001	0.101	0.220	0.011
	X9	X10	X11	X12	X13	X14	X15	X16
q-value	0.830 **	0.082	0.312 **	0.046	0.208 **	0.385 **	0.096	0.826 **
p-value	0.000	0.312	0.000	0.747	0.005	0.000	0.214	0.000

Note: * indicates significant at the 5% level; ** indicates significant at the 1% level.

). The results of factor detection, ranked by their explanatory power (q-value), are as follows: X₉ (0.830) > X₁₆ (0.826) > X₁₄ (0.385) > X₁₁ (0.312) > X₅ (0.255) > X₁₃ (0.208) > X₁ (0.200) > X₈ (0.162). This indicates that the density of the road network (X₉) and the annual average concentration of PM2.5 (X₁₆) exhibited the strongest statistical associations with the spatial heterogeneity of RR.

To further elucidate the interactions among driving factors, interaction detection was performed (Figure 7). More than 65% of the interacting factor pairs exhibited nonlinearly enhanced relationships, while the remainder showed two-factor enhancement. Key interacting factor pairs include X₁₄ ∩ X₁₆, X₉ ∩ X₁₄, and X₅ ∩ X₁₆. These results suggest a synergistic rather than purely additive driving mechanism behind the spatial heterogeneity of RR, indicating that it arises primarily from the complex interplay among multiple factors rather than from any single factor alone.

3.3.2. Determination of Key Drivers and Regression Model Comparison

To verify a statistically significant relationship with RR, Pearson correlation analysis was conducted on the eight previously identified significant factors. The analysis revealed that the correlation coefficient for X₅ was not statistically significant; consequently, this factor was excluded. To mitigate potential bias from multicollinearity among the remaining drivers, variance inflation factor (VIF) and tolerance tests were performed (

Table 6. Results of the correlation and multicollinearity tests of the driving factors.

Driving Factor	Correlation Test	Multicollinearity Test
	Pearson Correlation Coefficient	Covariance Statistics Before Exclusion		Covariance Statistics After Elimination
		p–Value	VIF	p–Value	VIF
X1	0.176 *	0.033 *	1.366	0.014 *	1.294
X5	−0.131	—	—	—	—
X8	0.309 **	0.161	1.337	—	—
X9	0.891 **	0.000 **	5.141	0.000 **	1.279
X11	0.452 **	0.100	1.504	0.046 *	1.330
X13	−0.253 **	0.002 **	1.232	0.002 **	1.191
X14	0.452 **	0.000 **	1.448	0.000 **	1.303
X16	−0.808 **	0.030 *	5.489	—	—

Note: * indicates significant at the 5% level; ** indicates significant at the 1% level. “—”signifies that the corresponding indicator was not included as a key driving factor in the final model.

). Ultimately, five driving factors passed all screening procedures: X₁, X₉, X₁₁, X₁₃, and X₁₄.

To identify the most suitable regression model for characterizing the driving mechanisms of RR, this study analyzed and compared the effects of various driving factors using classical ordinary least squares (OLS), GWR, and GTWR.

The Model comparison focused on two pivotal dimensions: explanatory power and specification adequacy. Explanatory power was quantified using a set of goodness-of-fit metrics, namely the coefficient of determination (R²), adjusted R², and the corrected Akaike Information Criterion (AICc). Specification adequacy is assessed by diagnosing whether significant spatial autocorrelation is present in the model residuals, as measured by the Global Moran’s I statistic. As presented in

Table 7. Performance comparison of regression models for assessing drivers of RR.

Model	R2	Adjusted R2	AICc	Residual Moran’s I	z-Score	p-Value
OLS	0.880	0.875	103.171	0.0523 **	15.595	0.000
GWR	0.955	0.953	63.837	−0.0084	−0.074	0.941
GTWR	0.976	0.975	33.756	−0.0081	0.019	0.985

Note: ** indicates significant at the 1% level.

, the GTWR model demonstrates superior performance on both criteria: it achieves the highest goodness-of-fit, and its residuals exhibited no significant spatial autocorrelation. Consequently, within this study, the GTWR model proved to be more applicable and robust than both the OLS and GWR models in capturing the spatiotemporal heterogeneity of RR driving mechanisms.

3.3.3. Temporal Evolution of Driving Mechanisms

GTWR regression analysis was conducted for each driving factor, and the temporal evolution of their regression coefficients was visualized using box plots (Figure 8). The sign of the coefficients reflects the direction of association between each driver and RR. The dispersion of the coefficients over time reflects the stability of each factor’s influence on RR.

The regression coefficient for the urban–rural income disparity (X₁) transitioned from positive to negative over time. In the initial stage, the income gap stimulated the outflow of rural labor. Some migrant workers remitted earnings to their rural households, creating a short-term “income window” that temporarily improved local economic conditions. Over time, however, the sustained loss of young and middle-aged laborers intensified the hollowing-out of rural communities. This prolonged decline in rural economic vitality ultimately exerted a negative influence on RR [55].

The regression coefficient for road network density (X₉) remained positive throughout the study period, despite demonstrating a fluctuating decline. Initially, road construction markedly enhanced local transportation accessibility, thereby facilitating the growth of specialty agriculture and rural tourism. However, constraints imposed by geological conditions led to considerably high costs for constructing and maintaining infrastructure. As highway network density reached a certain threshold, its marginal utility began to diminish, resulting in a gradually weakening contribution to RR.

The regression coefficient for the total agricultural machinery power per unit area (X₁₁) showed a trend of gradually declining positive values, with a tendency to eventually turn negative. The initial enhancement in agricultural mechanization significantly boosted production efficiency by substituting for human and animal labor [56]. However, as mechanization expanded, it exacerbated issues such as soil compaction and increased the risk of soil erosion. Concurrently, rising carbon emissions from machinery operations have collectively diminished the initially positive effects and may even lead to negative impacts on RR [57].

The regression coefficient for cropping structure (X₁₃) was generally negative but has exhibited a gradual increase since 2018, showing signs of turning positive. Within the KRBs, inherently low soil fertility and poor water retention capacity result in low production efficiency for food crops. Consequently, many farmers shifted to high-yield cash crops, which raised short-term income but led to a decline in regional grain production and a lower food self-sufficiency rate. In recent years, policy support aimed at boosting grain production, coupled with advancements in agricultural mechanization, has improved planting conditions and facilitated the scaling of food crop cultivation. These developments have enhanced the stability of the agricultural industry chain and strengthened the risk resistance capacity of the rural economy [58].

The regression coefficient for fractional vegetation coverage (X₁₄) transitioned from positive to negative values over the study period. In the initial phase, vegetation restoration effectively mitigated desertification in rocky areas and strengthened the region’s capacity for disaster prevention. However, as vegetation coverage continued to expand, it began to encroach upon agricultural production space, particularly in areas with shallow soil layers where tree roots disrupted soil structure and intensified soil erosion. Furthermore, the long investment return period associated with forestry projects made it difficult to offset losses in agricultural productivity, ultimately leading to a shift from a positive to a negative net effect on RR.

3.3.4. Spatial Heterogeneity of Driving Mechanisms

To examine the spatial heterogeneity of the impacts of driving factors, the average values of the regression coefficients across regions and study years were calculated, resulting in a spatial distribution map (Figure 9).

The impact of the urban–rural income disparity (X₁) on RR demonstrated a combination of positive and negative effects. In the central region, the strong siphoning effect from Guiyang—the provincial capital—has led to sustained outflows of crucial resources such as capital and talent, thereby constraining rural development. Conversely, counties in the peripheral areas are less susceptible to such urban siphoning effects, enabling them to retain more resources and maintain comparatively higher levels of RR.

The density of the road network (X₉) greatly enhanced RR. Weining County, located in Bijie City, is situated at the intersection of three provinces and serves as a vital transportation hub. A well-developed highway network mitigates barriers posed by karst topography. By leveraging the advantages of a transportation hub, economic connections and the flow of resources with surrounding areas were strengthened.

The impact of total agricultural machinery power per unit area (X₁₁) on RR demonstrated a dual nature. In areas characterized by highly fragmented arable land, the utilization of large-scale agricultural machinery is significantly constrained, thereby limiting potential gains in production efficiency. Conversely, regions with relatively favorable terrain conditions experience more substantial benefits from mechanization. In these areas, land consolidation initiatives have improved field flatness and connectivity, enabling a greater proportion of plots to accommodate mechanized operations and consequently enhancing agricultural productivity.

Cropping structure (X₁₃) generally exerted a negative influence on RR. Regions characterized by rugged terrain, severe rocky desertification, and highly fragmented arable land face constraints on large-scale and intensive agricultural development, often resulting in a monotonous planting structure. This lack of diversification makes the local economy highly susceptible to natural disasters and market fluctuations. Consequently, any significant disruption to primary crops can lead to a substantial downturn in the rural economy.

Fractional vegetation coverage (X₁₄) generally exerted a positive influence on RR across all counties, with particularly significant effects observed in ecologically fragile zones. For instance, in Shibing County, higher vegetation coverage helps mitigate soil erosion, maintain ecological balance, and reduce hazards associated with rocky desertification. By contrast, the effect of vegetation coverage is less pronounced in counties dominated by extensive agricultural production, where its role in supporting resilience remains relatively limited.

4. Discussion

4.1. Driving Mechanisms of RR in the Wujiang River Basin

The unique geological and ecological conditions of the Wujiang River Basin dictate that RR does not result from the isolated action of single factors. Instead, it is a coupled process wherein multiple driving factors adapt over time and interactively shape one another across spatial scales. Economic factors provide capital and resource support, while livelihood factors furnish service guarantees; together, they enable productive activities. The implementation of production factors affects rural economic vitality and the optimization of living services directly, while influencing the state of the ecosystem indirectly. In turn, dynamic changes in the ecological base feed back into economic inputs and production decisions. Through a closed loop of “support–feedback–constraint”, various factors form synergistic or balancing relationships. Their complex interactive effects shape the spatiotemporal differentiation pattern of RR.

4.1.1. Temporal Non-Stationarity and Threshold Effects of Driving Mechanisms

The period from 2010 to 2014 marked the initial phase of economic transformation and regional ecological restoration in the KRB. Infrastructure development and rocky desertification control advanced simultaneously. Road network density, the strongest explanatory factor, broke down transportation barriers, promoting specialty agriculture and rural tourism, thereby injecting initial momentum into resilience enhancement. Concurrently, labor outmigration triggered by the urban–rural income disparity brought remittance income in the short term, with both forces contributing to rural economic growth. This phenomenon validates the role of rural areas in supporting urban development through resource provision during the early [59]. Rocky desertification control promoted vegetation recovery, effectively mitigating soil erosion and risks, and thus consolidating the ecological foundation. However, constrained by fragmented karst farmland, the application of large agricultural machinery was limited, hindering the scalable advancement of agricultural mechanization. The market channels activated by road networks were not fully developed, leaving cash crops without stable sales outlets. Farmers still rely mainly on traditional grain crops, resulting in low production efficiency. Consequently, the potential effects of production-related factors were not fully realized during this period.

During 2014–2018, with the advancement of regional agricultural modernization and stronger market linkages, land consolidation and industrial structure optimization became developmental priorities. The implementation of land consolidation projects created conditions for the wider adoption of agricultural mechanization. Agricultural machinery power, by replacing human and animal labor, significantly boosted production efficiency, and its positive effect remained prominent [60,61]. Leveraging both market access facilitated by road networks and policy guidance, the cropping structure shifted to a diversified system integrating staple grains with cash crops, thereby mitigating their previously negative impact [62]. However, road construction in KRB faced high costs and significant engineering challenges, and over-reliance on infrastructure expansion to stimulate growth ultimately led to diminishing marginal returns from increased road network density. This finding aligns with Ayogu’s research on infrastructure development in ecologically fragile regions of Africa and with Ke’s analysis of the dynamic effects of transportation infrastructure in China. Ayogu notes that the impact of infrastructure on economic development is not absolute but highly context-dependent; pursuing quantitative expansion of hardware merely fails to achieve expected benefits and may lead to resource waste due to poor adaptability [63]. Ke’s empirical study in China’s upper-middle-income stage confirms that once infrastructure development passes a quantitative threshold, further expansion in metrics such as road mileage no longer drives growth significantly, instead, may cause an imbalance between investment and benefit if quality improvements and structural optimization are neglected [64].

From 2018 to 2022, as ecological conservation and urban–rural integration entered a new phase, pressures on ecological carrying capacity and imbalances in resource mobility within the KRB became more pronounced. Increasing vegetation coverage began to encroach on agricultural production space. In areas with shallow soil layers, tree roots damage soil structure, reducing the area suitable for mechanized operations [65,66]. Consequently, the positive effect of agricultural machinery peaked and then weakened, while the impact of vegetation coverage turned negative. This ecological threshold effect finds a cross-regional parallel in the research by Chen et al. on the Loess Plateau. Their study confirms the existence of a sustainable threshold for vegetation coverage between 53% and 65%, beyond which over-restoration conflicts spatially with agricultural production, validating the understanding that ecological carrying capacity has an objective upper limit [67]. Simultaneously, the urban–rural income gap and the improved road network formed a siphon effect. Core cities attracted young rural labor, exacerbating rural hollowing. As a result, the industrial stimulation effect of road networks weakened, and the urban–rural income gap exhibited a negative impact. This shift aligns with findings by Liu and Wu [68,69], challenging the conventional notion of a simple linear relationship between income disparity and rural decline [70].

These phenomena reflect the spatiotemporal non-stationarity in the effects of driving factors on RR. They indicate that in ecologically fragile areas, the effectiveness of policy interventions and developmental inputs critically depends on specific regional conditions, developmental stages, and baseline ecological constraints. When the intensity of human activities or environmental pressures surpasses the intrinsic carrying capacity, the impact of factors may undergo a shift, demonstrating strong nonlinearities and possible threshold effects. This phenomenon has been observed across multiple ecosystems globally [71,72,73,74,75].

4.1.2. Spatial Heterogeneity of Driving Mechanisms and the Role of Spatial Power

In the spatial dimension, the impacts of driving factors exhibit pronounced regional heterogeneity.

In central counties adjacent to the provincial capital, Guiyang, the asymmetry of spatial power relations, manifested through institutionalized resource allocation, continuously siphoned labor, capital, and policy resources from surrounding areas. Here, the urban–rural income disparity (X₁) concretely reflects such spatial deprivation, generally resulting in reduced RR [76]. In contrast, peripheral counties were better able to activate local ecological and cultural resources. By translating potential developmental disadvantages into endogenous drivers, these counties achieve higher levels of RR. This suggests that the spatial differentiation of RR arose from the interplay between a region’s position within broader power structures and its local response capacity. Notably, these findings challenge traditional core–periphery theory by demonstrating that peripheral regions are not necessarily doomed to passively accept disadvantage. Through targeted policy support and specialized industrial development, such areas can achieve leapfrog development—consistent with the insights of Kurikka and Leick [77,78].

The positive effect of density of the road network (X₉) is most pronounced at transportation hubs, underscoring the role of locational advantages in promoting development, consistent with Chen’s findings [79]. Similarly, the impact of total agricultural machinery power per unit area (X₁₁) exhibits both positive and negative effects across different regions, highlighting the critical importance of developing agricultural mechanization in a location-specific approach [80]. Furthermore, the varying impacts of cropping structure (X₁₃) and fractional vegetation coverage (X₁₄) across regions demonstrate that effective strategies to enhance RR must be tailored to local resource endowments through differentiated development approaches, consistent with the study by Geng [81].

This study reveals a deeper mechanism: the effects of driving factors are shaped not only by their intrinsic attributes but also by their embeddedness within regional spatial power structures. Interpreting RR as a dynamic outcome resulting from the interplay between a region’s positional advantage within the regional power network and its local response capacity breaks from the traditional perception of space as a static container. It reveals the agentic character of geographical space as both a medium for the operation of power and a carrier of social processes.

4.2. Policy Recommendations

These findings necessitate a rethinking of governance frameworks in ecologically fragile areas. Conventional management paradigms, which rely on uniform standards, prove inadequate in addressing the dynamic complexities of these systems. Policy design needs to achieve four key shifts:

(1)

Shift toward differentiated regulatory mechanisms based on ecological carrying capacity assessments. In core karst zones with severe rocky desertification and high ecological fragility, priority should be given to planting pioneer tree species that tolerate poor soils, while controlling inefficient anthropogenic disturbances. In peripheral zones with milder desertification and more concentrated farmland, focus on maintaining vegetation coverage within an optimal range This approach strengthens the ecological buffer for core zones and safeguards sustainable agricultural space.

(2)

Shift from single-objective optimization to multi-objective synergy, seeking systemic optima by balancing economic efficiency and ecological sustainability. In core zones, flexible ecological ranger positions could be established, they would be responsible for vegetation patrol and soil erosion control. In peripheral zones, contiguous farmland should support the scaled cultivation of specialty crops. Leveraging improved road networks, processing enterprises can meet labor needs by drawing workers from core zones.

(3)

Shift from end-of-pipe intervention to whole-process supervision by establishing a differentiated early-warning and response system. In core zones, focus on the dual constraints of ecological thresholds and factor outflow, tracking vegetation coverage and labor retention rates. If vegetation coverage approaches the lower threshold, a tiered emergency restoration protocol should be activated. A sustained increase in outmigration rates should trigger parallel incentives for ecological ranger positions to prevent a vicious cycle between conservation and labor loss. In peripheral zones, the road investment per unit of agricultural output should be adopted as a key monitoring indicator. An increase over two quarters should prompt a shift in infrastructure planning from expanding mileage to improving supporting facilities.

(4)

Strengthen regional coordination and linkage, and improve cross-regional collaborative safeguarding mechanisms. By innovating ecological compensation and horizontal fiscal transfer mechanisms, the positive externalities generated through ecological protection can be quantified scientifically and compensated fairly. This helps address imbalances in development rights caused by asymmetric functional divisions. Counties near core cities should focus on enhancing resource feedback mechanisms and strengthening industrial support and public services. Counties farther from core cities, efforts should increase their capacity to create ecological products and promote the transformation of unique local resources into innovative, sustainable development pathways.

4.3. Research Advantages and Limitations

This study addresses a gap in the literature on RR, which often focuses on plain areas or general ecological zones, by centering on the distinctive context of the Wujiang River Basin—a typical karst basin. It constructs a comprehensive RR evaluation index system encompassing three dimensions: agricultural production, residential living, and the ecological environment. By selecting and refining targeted indicators, it mitigates the poor adaptability of generic evaluation systems to the unique topography and ecological characteristics of karst regions. Methodologically, the study integrates the OPGD and GTWR models. This combined approach overcomes the limitations of the OLS model’s assumption of global homogeneity and the traditional GWR model’s inability to capture temporal non-stationarity in driving mechanisms, thereby providing a more regionally tailored empirical framework and methodological support for RR research in karst basins.

However, this study has certain limitations. First, despite its multidimensional design, the evaluation system does not sufficiently incorporate socio-cultural indicators of rurality. Factors that are difficult to quantify, such as the preservation of traditional culture, were not fully accounted for. Secondly, the data primarily relies on statistical yearbooks and monitoring reports, lacking micro-level survey data (e.g., on household farming decisions). The update frequency and precision of some ecological indicators also constrained the analytical accuracy. Finally, although the GTWR model effectively captures spatiotemporal non-stationarity in parameters, it remains a form of local linear regression. It does not fully circumvent biases inherent to linear assumptions and may not adequately capture complex nonlinear effects.

Future research should aim to incorporate socio-cultural indicators to refine the resilience evaluation framework. Combining macro-statistical data with micro-level household surveys would enhance the comprehensiveness and precision of mechanism analysis. Furthermore, exploring machine learning or nonlinear regression models could help address the limitations of the GTWR model’s linear assumptions and better capture nonlinear interactions among driving factors.

5. Conclusions

This study has investigated the spatiotemporal evolution and driving mechanisms of RR in the ecologically fragile Wujiang River Basin, a typical KRB. By constructing an integrated analytical framework combining the OPGD and GTWR models, a systematic analysis of long-term time series data (2010–2022) was conducted. The main findings are as follows:

(1)

From 2010 to 2022, RR levels showed a significant upward trend across the Wujiang River Basin, alongside a marked narrowing of inter-county disparities. This indicates improved regional coordination and more balanced development. Spatial autocorrelation analysis revealed a significant positive spatial dependence of RR. The observed clustering patterns underwent dynamic evolution, shaped by the combined effects of policy interventions and environmental constraints.

(2)

The driving mechanisms of RR exhibited pronounced spatiotemporal non-stationarity and complex nonlinearity. Key drivers identified include the urban–rural income disparity (X₁), density of the road network (X₉), total agricultural machinery power per unit area (X₁₁), cropping structure (X₁₃), and fractional vegetation coverage (X₁₄). The GTWR results demonstrated that the effects of these drivers are subject to significant spatiotemporal non-stationarity. Notably, the direction of influence of factors such as the urban–rural income gap and vegetation coverage has reversed over time. This shift indicates that, in ecologically fragile areas, the efficacy of interventions is constrained by ecological thresholds. Furthermore, the heterogeneity in the impacts of driving factors is closely linked to regional spatial power structures, revealing that geographic space acts not merely as a passive container for development but as an active medium shaping resilience processes.

These findings provide a scientific basis for policymakers to formulate tailored strategies to enhance regional RR and promote sustainable development. The methodology and empirical insights presented here can serve as a reference for research and practice concerning RR in other similar KRBs and environmentally sensitive regions globally.