Threshold Effects of Biodiversity on Ecological Resilience: Evidence from Guangdong’s Prefecture-Level Cities

Abstract

Understanding interactions between ecological resilience and biodiversity is critical for sustainable ecosystems and coordinated regional development. This study examines prefecture-level cities in Guangdong Province—characterized by diverse ecological conditions and rapid urbanization—to explore how ecological systems respond to biodiversity dynamics. We construct an ecological resilience framework based on resistance–adaptability–recoverability, quantify biodiversity using species occurrence data from the Global Biodiversity Information Facility, and apply a panel threshold model to detect nonlinear couplings. To identify key drivers of resilience, we employ XGBoost and SHAP analyses for interpretable machine learning insights. Results show clear threshold behavior: ecological resilience is weak or negative at low biodiversity and improves once biodiversity exceeds critical levels; in 2015 and 2020, thresholds were approximately 99.73 and 232.01 with a significant high-biodiversity effect. Machine learning results align with the threshold findings and indicate forest coverage ratio is the dominant driver of ecological resilience across years. The integrated findings highlight pronounced spatial heterogeneity in ecological resilience and identify critical biodiversity thresholds influencing ecosystem stability, providing targeted evidence for biodiversity conservation and resilience-oriented management. This study advances understanding of nonlinear ecological–biodiversity interactions and offers practical guidance for strengthening ecological security in rapidly developing regions.

1. Introduction

Climate change intensifies disturbances to ecosystems worldwide, posing growing challenges to biodiversity conservation and long-term ecological resilience. The Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report shows that climate change has already caused substantial and partly irreversible losses in terrestrial and freshwater ecosystems, with nearly half of assessed species shifting their ranges to higher latitudes or elevations. Future projections indicate that Asia, including southern China, will experience greater temperature extremes, rainfall variability, and drought events, further threatening biodiversity and ecosystem stability. These findings underscore the urgent need to strengthen ecological resilience as a foundation for sustaining ecosystem functions and mitigating biodiversity loss under medium- and long-term climatic change.

Biodiversity (BD), as a cornerstone of sustainable development, not only underpins ecosystem functioning but also profoundly influences ecological resilience [1,2]. Understanding the interactions between BD and ecological resilience is therefore essential for advancing coordinated ecological management and achieving long-term sustainability [3,4,5]. This study focuses on cities across Guangdong Province, selected as an empirical context due to their rapid urbanization, diverse ecological conditions ranging from coastal plains to mountainous areas, and the availability of fine-scale BD data. These characteristics make Guangdong an ideal setting to examine how urban ecosystems respond to variations in BD and to explore spatial heterogeneity in resilience patterns.

Ecological resilience (ER), first conceptualized by Holling [6], refers to an ecosystem’s capacity to maintain functional stability, adapt to disturbances, and recover from disruptions. Unlike traditional equilibrium-based approaches, resilience emphasizes dynamic system evolution, highlighting the optimization of ecosystem services, efficient resource utilization, and coordinated human–environment interactions. Within Guangdong, cities differ substantially in resource endowment, land-use intensity, and developmental stage, producing gradient patterns in ER and BD interactions (Figure 1 and Figure 2). For instance, Shenzhen faces intense ecological pressures due to dense urbanization and limited natural habitats, whereas Zhaoqing retains a relatively intact ecological base but exhibits weaker resilience amid rapid land-use changes. Ongoing BD conservation policies, habitat restoration efforts, and ecological red-line planning continue to shape urban ecosystems in Guangdong, resulting in varied capacities to resist, adapt, and recover from environmental disturbances [7,8].

A review of existing literature highlights three key aspects of research on ER and BD. First, evaluation frameworks either adopt process-oriented approaches grounded in resilience theory—such as “pressure–state–response” and “risk–connectivity–potential”—or construct composite structure–function indices tailored to regional spatial characteristics [9,10,11]. Second, methodological approaches range from geographic detectors and coupling coordination models to regression analyses, facilitating the study of synergies, trade-offs, and complex nonlinear interactions [12]. Third, most studies have focused on national, regional, or ecosystem-level scales, with relatively few comparative analyses across cities within the same province, limiting the ability to inform localized ER and BD policies.

Despite these advances, several gaps remain. Most existing studies focus on broad-scale ER–BD coupling, overlooking gradient differences among cities. Moreover, limited research has examined how ER varies under different BD scenarios at the provincial level. To address these gaps, this study takes prefecture-level cities across Guangdong Province as empirical units, constructs a multidimensional ER framework encompassing resistance, adaptability, and recoverability, and integrates coupling coordination analysis with interpretable machine learning. In addition, the study not only investigates the relationship between ER and BD but also recognizes the potential limitations of observation-based biodiversity data. Given that observation-based biodiversity data are strongly influenced by research capacity, economic development, and population distribution, the study further discusses how observation bias may affect model interpretability and threshold identification. This reflection provides valuable methodological insights for the effective use of open biodiversity data in regional ecological assessments.

This paper is organized into five sections. The first section introduces the research background, objectives, and significance. The second section presents a comprehensive literature review on ER and BD. The third section details the research methodology, including index construction, threshold modeling and machine learning interpretation. The fourth section reports empirical results, highlighting ER spatial heterogeneity, nonlinear interactions, and key drivers. The fifth section provides discussion and conclusions, offering both theoretical insights and practical policy recommendations for strengthening biodiversity conservation and enhancing ecological resilience in rapidly urbanizing regions.

2. Literature Review

Based on the systematically screened studies, research on the relationship between ER and BD has developed along multiple trajectories, reflecting both theoretical advances and methodological innovations. To ensure transparency and reproducibility, the review follows PRISMA 2020 principles and systematically searches leading academic databases for studies published between 1995 and 2025. The review searches Web of Science, Scopus, and Google Scholar for English-language, peer-reviewed publications using combinations of the terms “ecological resilience,” “biodiversity,” and “urban ecosystem.” Studies are included if they explicitly examined ER–BD relationships via empirical or conceptual approaches at ecosystem, regional, or urban scales; conference abstracts and reports lacking methodological detail are excluded.

BD is widely recognized as a fundamental driver of ecosystem stability, with higher species richness and functional diversity enhancing ecosystems’ capacity to withstand disturbances, adapt to changing conditions, and recover from shocks [13,14,15]. ER, first conceptualized by Holling [6], has evolved into a multi-dimensional framework emphasizing resistance, adaptability, and recoverability. Within this framework, BD plays a pivotal role in stabilizing ecological processes and sustaining ecosystem services, bridging ecological theory with practical conservation strategies [16,17,18].

Through thematic coding of the included records, two major perspectives dominate. The first emphasizes BD as an intrinsic component of resilience, highlighting mechanisms such as redundancy, complementarity, and functional diversity that enhance system stability [19,20,21]. The second perspective treats BD as an external driver of resilience outcomes, where changes in species richness, composition, or habitat structure influence ecosystem thresholds and tipping points [22,23]. Both perspectives underscore the importance of understanding how BD loss or conservation reshapes resilience under anthropogenic pressures, including urbanization, land-use change, and climate variability. Evidence is synthesized thematically, drawing on convergences and divergences across indicators, spatial scales, and study designs.

Methodologically, scholars have adopted a diverse array of tools to evaluate ER–BD interactions. Early studies often relied on qualitative assessments of ecosystem stability, whereas more recent research integrates quantitative indices, such as species richness, Shannon-Wiener diversity, and functional trait metrics derived from global repositories like GBIF [24,25,26]. Analytical approaches include geographic detectors, coupling coordination models, and regression analyses, which enable the examination of synergies, trade-offs, and complex interactions between ER and BD [27]. Recent advances incorporate nonlinear models, such as threshold regression and structural equation modeling, as well as machine learning algorithms, providing deeper insights into critical drivers of resilience under varying BD conditions [28,29,30].

In terms of spatial scale, most ER–BD studies focus on global, national, or ecosystem-specific levels, such as forests, wetlands, or coastal zones. Comparative studies at regional or provincial scales, particularly across multiple cities, remain limited, constraining the development of localized conservation and resilience strategies [31]. Similarly, temporal analyses emphasize long-term ecological change, with fewer studies exploring the dynamic evolution of resilience in rapidly urbanizing contexts, where BD pressures and conservation measures interact in complex ways [32,33,34].

Despite these advances, several research gaps remain. First, spatial heterogeneity at the city level is often underexplored, yet is crucial for designing targeted management interventions [35]. Second, nonlinear interactions and threshold effects are rarely examined, despite their importance in ecosystems that exhibit tipping points [36]. Third, although global BD data repositories such as GBIF offer unprecedented opportunities for comparative analysis, few studies leverage these data to assess resilience across urban systems [37,38,39]. Addressing these gaps requires integrating high-resolution BD metrics into resilience frameworks, applying nonlinear models to detect critical thresholds, and employing interpretable machine learning approaches to identify key drivers of resilience.

To address the aforementioned gaps, this study adopts a city-level perspective, integrating ER dimensions, BD metrics and socioeconomic factors to construct a comprehensive analytical framework. The framework is guided by three principles: (1) multidimensional resilience assessment, measuring resistance, adaptability, and recoverability to capture ecosystem functionality; (2) threshold analysis, using nonlinear interaction modeling to identify critical biodiversity levels required for significant resilience gains; and (3) driver identification, employing machine learning (XGBoost + SHAP) to quantify the relative importance of BD, forest coverage, and economic factors in shaping resilience patterns.

This framework explicitly accounts for spatial heterogeneity among cities and temporal dynamics over 2010–2020, enabling both cross-sectional comparisons and longitudinal analyses. By integrating these methodological approaches, the study overcomes the limitations of previous research that focused on broad-scale or single-dimension assessments and provides a structured pathway for evaluating threshold effects, synergies, and trade-offs in urban ecosystems (Figure 3).

This study makes several notable contributions to advancing understanding of the ER, BD and addressing persistent gaps in the literature. First, it integrates high-resolution, city-level BD data from GBIF, enabling a spatially explicit quantification of ecological diversity across urban and peri-urban landscapes. This approach overcomes limitations of coarse-scale or aggregated BD metrics, allowing for a more precise examination of how species richness and distribution patterns influence ecosystem functionality. Second, the study employs panel threshold modeling to uncover nonlinear coupling effects, revealing critical BD thresholds that determine whether ER is enhanced. This insight moves beyond conventional linear assessments, highlighting the existence of ecological tipping points where incremental BD gains yield minimal benefits unless thresholds are surpassed. Third, by leveraging interpretable machine learning techniques (XGBoost + SHAP), the study identifies the dominant ecological and socio-economic drivers of resilience, including forest coverage ratio, BD, Gross Domestic Product (GDP), urban built-up area ratio, total area of nature reserves and extreme climate events. This approach bridges traditional ecological modeling with modern data-driven methods, offering a transparent mechanism to assess the relative importance of multiple interacting factors.

Collectively, these contributions provide both theoretical and practical value. The framework enriches understanding of ER–BD interactions under rapid urbanization, elucidates the nonlinear dynamics that govern ecosystem responses, and informs evidence-based urban ecological planning and conservation strategies. By linking high-resolution BD metrics, threshold-based analyses, and interpretable driver assessment, the study establishes a robust foundation for policy-relevant interventions aimed at sustaining resilient urban ecosystems in the face of anthropogenic pressures and spatial heterogeneity. To operationalize this framework and address the identified gaps, we next employ an ecological resilience evaluation model, a panel threshold model, and XGBoost–SHAP to detect thresholds and the key drivers of ER.

3. Methodology

3.1. Study Context

Guangdong Province, located in southern China, provides an ideal context for examining ER and BD dynamics due to its diverse topography, rapid urbanization, and rich species diversity (Figure 4). The province encompasses low-elevation plains and densely urbanized areas, such as the Pearl River Delta, as well as high-elevation, forested mountainous regions, creating strong spatial gradients in ecosystem structure and function. Rapid economic development and land-use changes have exerted heterogeneous pressures on ecosystems, while ongoing conservation policies—including forest protection, habitat restoration, and ecological red-line planning—shape local resilience patterns. This combination of natural heterogeneity, anthropogenic influence, and available high-resolution BD data enables a nuanced analysis of how ER varies across space and responds to BD thresholds, providing insights relevant to both regional management and broader urban sustainability challenges.

3.2. Data Processing

Land-use remote sensing image is preprocessed through clipping, atmospheric correction, and geometric correction. To ensure consistency across datasets with differing spatial resolutions, ArcGIS Pro 3.0.0 tools such as clip, mask extraction, and raster resampling are applied. All datasets are unified to the WGS 1984_UTM Zone_49N coordinate system with a 30 m spatial resolution.

BD data are obtained from the Global Biodiversity Information Facility (GBIF) as species-occurrence records for the study region. Data processing comprises removal of duplicate entries, validation of geographic coordinates and species-level identifications, and aggregation at the prefecture-city and year level. Because observation effort varies across cities and years, an effort-aware sampling normalization is implemented prior to analysis: city–year summaries of observation effort, together with year fixed effects to capture platform-wide shifts, are used to align all cities to a common, year-specific median effort baseline, after which GBIF-based BD is rescaled with a quadratic penalty on the effort gap. This bidirectional adjustment proportionally downweights high-effort cities and modestly upweights low-effort cities without altering the BD scale, thereby reducing inflation or deflation due to sampling intensity and yielding BD values more commensurate with ecological resilience for cross-city comparison (Figure 5). The resulting BD metrics are then spatially aligned with land-use layers to support integrative analysis.

3.3. Ecological Resilience Evaluation Model

Resistance (P): Resistance refers to the ability of an ecosystem to maintain its structure and functions in the face of external disturbances or pressures and to quickly recover to its initial state. The richer the BD and the more stable the structure, the stronger the ecological service functions and the higher the resistance. Thus, ecosystem service value (ESV) is used as a proxy for resistance. Following Costanza et al.’s classification of ecosystem service functions and adjusted to Guangdong’s land-use characteristics, value coefficients are assigned to different land-use types [40]. The ecosystem service value of construction land is set to zero. Based on Xie Gaodi’s research, one-seventh of the unit crop production value was used as the equivalence factor for calculating ecosystem service value [41]. The formula is:

$$P=ESV=\sum A_k\times V{C}_{kj}$$

(1)

where ESV is the ecosystem service value, Aₖ is the area of land-use type k, VCₖⱼ is the value coefficient of land-use type k for service j, and the value coefficients for the land-use types in Guangdong are sourced from Ye and Dong [42].

Adaptability (A) represents the capacity of an ecosystem to maintain the integrity of its structure and function through self-organizing regulation under external disturbances. When natural disasters or human activities disrupt landscape equilibrium, the system requires a period of adjustment to regain stability. To quantify the structural stability of ecosystems, this study incorporates indicators closely related to landscape heterogeneity and landscape connectivity, which are key determinants of landscape stability. Landscape indices are employed as quantitative tools due to their ability to precisely characterize spatial patterns, reveal structure–function coupling relationships, and monitor dynamic system evolution. Given that landscape heterogeneity and connectivity play equally important roles in maintaining landscape stability, this study assumes equal weighting between them [43]. Specifically,

$$A=0.25\ast SDI+0.25\ast AWMPFD+0.5\ast LFI$$

(2)

where SDI is the Shannon’s Diversity Index, AWMPFD is the Area-Weighted Mean Patch Fractal Dimension, LFI is the Landscape Fragmentation Index.

Resilience (R) refers to the ability of an ecosystem to return to a stable state and maintain its essential characteristics after being disturbed by external factors such as extreme climatic events or human activities. The calculation formula is as follows:

$$R=\sum _{k=1}^n{A}_k\times R{C}_k$$

(3)

where R represents the ER of the ecosystem, A_k denotes the proportion of land-use type k, and RC_k is the resilience coefficient of land-use type k, with coefficient values referenced from Peng [44].

The final calculation formula for ER is as follows:

$$ER=\sqrt[3]{P\times A\times R}$$

(4)

where ER represents the comprehensive ecological resilience level, P denotes resistance, A denotes adaptability, and R denotes recovery.

3.4. Panel Threshold Model

To examine potential nonlinear relationships and threshold effects between biodiversity dynamics (BD_it) and ecological resilience (ER_it), this study employs the panel threshold model proposed by [45] and reports formal inference for the threshold—bootstrap p-values for the no-threshold test and 95% confidence intervals for γ. This model is particularly suited for identifying critical thresholds in panel data, where the effect of an explanatory variable on the dependent variable may shift discontinuously once the threshold is crossed:

$${ER}_{it}={\mu }_i+{\beta }_1{BD}_{it}\times I\left({BD}_{it}\le \gamma \right)+{\beta }_2{BD}_{it}\times I({BD}_{it}>\gamma )+\delta ’X_{it}+{ϵ}_{it}$$

(5)

where

ER_it is ecological resilience of city I at time t.

BD_it is the core explanatory variable, representing biodiversity dynamics

I(·) is an indicator function: I(BD_it ≤ γ) = 1 if BD_it ≤ γ, and 0 otherwise; similarly, I (BD_it > γ) = 1 if BD_it > γ, and 0 otherwise;

β₁ and β₂ are the regression coefficients of BDit in the low and high regimes, respectively, capturing the differential effects of BD on ER across the threshold;

X_it represents a set of control variables, including Forest Coverage Ratio, GDP, Urban Built-up Area Ratio, Total Area of Nature Reserves, Extreme Climate Events;

$\delta$ Coefficient vector of control variables;

ϵ_it is the error term, represents unobserved factors and random shocks affecting ER that are not explained by the model variables.

3.5. Machine Learning Interpretation (XGBoost + SHAP)

To further explore which variables most strongly influence ER, an XGBoost regression model is trained with BD, Forest Coverage Ratio, GDP, Urban Built-up Area Ratio, Total Area of Nature Reserves and Extreme Climate Events as predictor variables.

SHAP (SHapley Additive exPlanations) is applied to interpret model outputs, quantifying each feature’s contribution to ER predictions. This approach provides a transparent and quantitative understanding of the relative importance of BD, Forest Coverage Ratio, GDP, Urban Built-up Area Ratio, Extreme Climate Events and Total Area of Nature Reserves variables, complementing the panel threshold analysis.

4. Results

4.1. Ecological Resilience Based on Resistance–Adaptability–Recoverability

Evaluation of ER in Guangdong from 2010 to 2020 reveals significant spatio-temporal differentiation across cities. The multidimensional assessment using resistance, adaptability, and recoverability (R–A–R) provides a comprehensive understanding of ecosystem stability and dynamics under varying environmental and land-use conditions.

Resistance, representing the ability of ecosystems to maintain structure and function under external disturbances, displays a clear core–periphery polarization (Figure 6). Low-resistance areas are predominantly located in low-elevation plains, whereas high-resistance areas are concentrated in mountainous forests. Quantitatively, the proportion of high-resistance areas decreased from 6.6% in 2010 to 4.0% in 2020, while upper-middle resistance areas declined from 22.3% to 17.6%, indicating a gradual decline in the resilience advantages of previously stable regions. Global Moran’s I is positive and highly significant at the prefectural scale (I = 0.167 in 2010, 0.156 in 2015, 0.191 in 2020; row-standardized k-nearest-neighbor weights; all p < 0.001), indicating persistent clustering with a modest uptick by 2020.

Adaptability, defined as the capacity of ecosystems to adjust and reorganize in response to disturbances, shows a pattern roughly inverse to that of resistance (Figure 7). The proportion of high-adaptability areas decreased from 6.4% in 2010 to 4.1% in 2020. Low-level adaptability areas also contracted slightly (from 52.3% to 50.4%), while middle-level areas increase. Global Moran’s I remains stable over time (I = 0.315 in 2010, 0.319 in 2015, 0.310 in 2020; row-standardized k-nearest-neighbor weights; all p < 0.001), pointing to durable pockets of high adaptability and limited dispersion.

Recoverability, reflecting the speed and extent to which ecosystems can return to their original state following disturbances, shows an overall declining trend across Guangdong (Figure 8). High-recoverability areas decreased from 17.6% in 2010 to 14.2% in 2020, while low-recoverability areas increased from 15.3% to 17.2%. Notably, high-recoverability zones remain concentrated in high-elevation, forest-dominated, and less-disturbed regions. Global Moran’s I for Recoverability is consistently high at the prefectural scale (I = 0.679 in 2010, 0.691 in 2015, 0.690 in 2020; row-standardized k-nearest-neighbor weights; all p < 0.001), indicating pronounced and persistent positive spatial autocorrelation.

Overall, the R–A–R evaluation reveals a shifting ecological landscape in Guangdong, characterized by decreasing high-resilience zones, consolidation toward moderate adaptability, and declining recoverability (

Table 1. Trends in Resistance, Adaptability, and Recoverability (2010–2020).

		Low		Lower-Middle		Middle		Upper-Middle		High
Resistance	2010	20.96%	↑ *	19.16%	↑	30.98%	↑	22.29%	↓	6.61%	↓
	2015	22.59%		18.72%		31.34%		20.94%		6.42%
	2020	21.81%		19.83%		36.71%		17.63%		4.02%
Adaptability	2010	52.28%	↓	22.26%	↑	13.73%	↓	5.36%	↑	6.36%	↓
	2015	52.09%		24.81%		13.71%		5.26%		4.12%
	2020	50.39%		26.41%		13.55%		5.60%		4.05%
Recoverability	2010	15.27%	↑	25.66%	↓	18.95%	↑	22.55%	↑	17.57%	↓
	2015	16.27%		24.90%		19.42%		23.08%		16.32%
	2020	17.21%		25.16%		19.77%		23.66%		14.20%

* An upward arrow (↑) indicates that resistance, adaptability, or recoverability increased from 2010 to 2020, while a downward arrow (↓) indicates that these indicators decreased over the same period.

).

Overall ER maps for 2010–2020 show strong spatial heterogeneity (Figure 9). By 2020, low and lower-middle resilience dominated highly urbanized and agriculturally intensive areas such as Foshan (67.79%), Dongguan (56.44%), Zhongshan (55.71%), and Zhanjiang (55.26%). High and upper-middle resilience were concentrated in northern, forest-rich cities like Qingyuan (21.11%), Heyuan (20.55%), and Zhaoqing (18.8%). Temporal trends reveal that, except for Chaozhou (+1.85%) and relatively stable resilience in Shantou, Shanwei, and Heyuan, most other cities experience continuous declines. Jiangmen (−8.20%) showed the steepest deterioration, followed by Foshan (−6.54%), Guangzhou (−5.98%), Shenzhen (−5.66%), Zhuhai (−5.66%), and Yangjiang (−5.74%). Across Guangdong, the average ER fell by 3.96% between 2010 and 2020. Global Moran’s I for ER indicates clear and strengthening spatial dependence at the prefectural scale: 0.193 (2010), 0.211 (2015), and 0.224 (2020) based on row-standardized k-nearest-neighbor weights (all p < 0.001).

4.2. Threshold Effects Between ER and BD

In 2010, the estimated BD threshold is 35.61 (95% CI: 5.91–96.85), and the Hansen test does not support a statistically significant threshold effect (threshold p = 0.82), suggesting that nonlinear BD–ER segmentation is weak and poorly identified at this stage. By 2015, the optimal BD threshold rises to 99.73 (95% CI: 99.73–100.89), and the threshold effect is statistically significant (threshold p = 0.03). Below the threshold, the BD coefficient is negative (r = 0.40), whereas above the threshold it turns slightly positive (r = 0.43). The threshold specification substantially improves the model fit (R² = 0.8706 compared with 0.5629 for the no-threshold model), pointing to a pronounced nonlinear structure in the BD–ER relationship by the mid-2010s. In 2020, the estimated BD threshold increases to 232.01 (95% CI: 232.01–233.19), and the threshold effect is statistically significant at the 5% level (threshold p = 0.05). The simple BD–ER correlations differ in sign, being strongly negative at lower BD levels (r = −0.72) and weakly positive at higher BD levels (r = 0.17), with the corresponding p-value of 0.6654 indicating that this positive association is not statistically significant. Although these marginal effects are modest in magnitude, the threshold specification attains a very high goodness of fit (R² = 0.9643 versus 0.8737 for the no-threshold model) (Figure 10 and Figure 11).

4.3. Machine Learning Interpretation of Driving Factors

An XGBoost regression model is developed using BD, forest coverage ratio, GDP, urban built-up area ratio, extreme climate events, and total area of nature reserves as predictor variables, with ER as the target variable at the prefecture-level city scale. After excluding outliers, 21 valid samples are retained. The model is trained with a learning rate of 0.1, maximum tree depth of 3, 100 boosting rounds, and a mean squared error loss function, with hyperparameters selected via grid search and cross-validation to balance predictive performance and overfitting. To further assess robustness given the small sample, leave-one-out cross-validation (LOOCV) is conducted, confirming consistent predictive performance across folds (Figure 12).

The XGBoost 3.1.0 analysis summarize both predictive performance and the relative importance of explanatory variables. LOOCV-predicted ER values are close to the 1:1 reference line when plotted against observed ER. Feature importance analyses, including permutation importance measures (Figure 13 and Figure 14), consistently identify the Area of Nature Reserves, Forest Coverage Ratio, and BD as the variables with the highest importance scores, while GDP and climate-related variables (Extreme Climate Events, Extreme High Temp Days) have importance values close to zero.

To investigate temporal dynamics, ER distributions and predictor importance are compared across 2010, 2015, and 2020. The Kruskal–Wallis H test reveals no significant differences in ER across the three years (H = 1.653, p = 0.438; Figure 15). In contrast, the Friedman test shows significant temporal variation in feature contributions (χ² = 16.571, p = 0.011; Figure 16), with the importance of forest coverage ratio increasing from 0.261 to 0.950, BD decreasing from 0.212 to 0.024, and urban built-up area ratio fluctuating over the decade.

5. Discussion

This study uncovers complex and nonlinear interactions between ecological resilience (ER) and biodiversity (BD) across Guangdong’s prefecture-level cities between 2010 and 2020. Evaluation along the three dimensions—Resistance, Adaptability, and Recoverability (R–A–R)—reveals that resilience has become increasingly uneven and dimension-specific. High levels of resistance and recoverability have contracted and are now largely confined to high-elevation, forest-dominated and relatively undisturbed areas, whereas high adaptability is concentrated in low-lying, heavily urbanized corridors where socio-ecological systems retain a greater capacity to reorganize in the face of disturbance. At the same time, lowland plains that have undergone intensive urbanization show marked losses in resistance and recoverability, indicating reduced ability to buffer and recover from shocks. Over the period 2010–2020, ER declined on average and became more spatially uneven, with many lowland cities shifting into low or lower-middle resilience classes. These patterns suggest that not all ecosystems, nor all dimensions of resilience, respond in the same way to anthropogenic pressures and point to emerging vulnerability hotspots in heavily modified urban landscapes where key resilience mechanisms may already be close to being overwhelmed. Similar spatial polarization of ecological resilience, with high-resilience “cores” and low-resilience “peripheries”, has been reported in studies of Chinese urban agglomerations and coastal regions, where rapid land-use transformation and production–living–ecological space conflicts tend to suppress resilience in intensively developed areas [46,47,48].

The threshold analysis provides further insight into the nonlinear role of BD in shaping ER. In 2010, the estimated BD threshold is imprecise and statistically non-significant, indicating limited evidence for a clear breakpoint in the ER–BD relationship. By 2015 and 2020, however, statistically significant thresholds emerge and shift toward higher BD levels, indicating that the association between BD and ER changes across the BD gradient. In 2015, a threshold around moderate-to-high BD levels separates a regime in which additional biodiversity is associated with lower resilience from one in which this adverse marginal effect is largely neutralized or slightly mitigated. By 2020, the estimated threshold moves to very high BD levels, and ER remains negatively related to BD on both sides of the threshold, although the magnitude of the negative effect is reduced above the threshold. Correlation analyses also show that the sign and strength of the ER–BD association vary between low- and high-BD ranges, with a strong negative correlation at low BD and a weak, statistically non-significant positive correlation at high BD in 2020. Taken together, these results indicate that the influence of BD on ER is nonlinear and regime-dependent: clear improvements in ER are not observed at low BD levels, and even where thresholds are statistically significant, the marginal effects are modest in size. This pattern is consistent with threshold responses documented in other ecosystems, where biodiversity or habitat loss produces abrupt changes in species richness, ecological condition, or ecosystem function once critical thresholds are crossed [49,50,51]. In those studies, thresholds often occurred at specific levels of forest cover or land-use intensity, whereas in the present work thresholds are expressed in terms of city-level BD under strong urbanization pressure, highlighting how the location and magnitude of thresholds depend on both ecological context and the choice of controlling variables. The comparatively weak and unstable threshold evidence in the early period is likely to reflect both ecological conditions and limitations of observation-based BD data in lower-capacity cities, where low observation effort may have prevented some systems from reaching or revealing their effective biodiversity levels. As data completeness and spatial coverage improve over time, the sharper thresholds observed in 2015 and 2020 may reflect a combination of genuine ecological change and reduced measurement bias, but they still call for cautious interpretation in data-poor regions.

The XGBoost 3.1.0 and SHAP 0.49.1 analyses complement the threshold results by clarifying the relative importance of different drivers. Across the three time points, forest coverage and the area of nature reserves consistently emerge as the most influential variables for explaining spatial variation in ER, followed by the urban built-up area ratio, while BD plays a more moderate and declining role in the models. In contrast, economic indicators (GDP) and climate-related variables (extreme climate events and extreme high temperature days) have importance values close to zero, indicating limited predictive contribution within the temporal and spatial scope of this study. This dominance of forest structure and protection status is consistent with landscape-ecology and connectivity research showing that forest amount and connectivity underpin biodiversity conservation, ecological processes, and resilience in forested and urbanizing regions [52,53,54]. Studies of urban forests in China similarly highlight that core habitat within urban forest patches is crucial for ecological functionality and resilience, even as recreational functions expand [55]. The strengthening importance of forest coverage over time, together with the decreasing contribution of BD, suggests that structural habitat attributes—such as the amount and continuity of forested and protected land—may be more robust predictors of ER than point-based species occurrence records in rapidly urbanizing, data-limited contexts. The rising importance of forest coverage and the persistently high Moran’s I values for recoverability also support the interpretation that forest-dominated regions act as structural “backbones” of resilience, while many lowland urban and agricultural systems show both lower ER levels and weaker contributions from BD. This interpretation is in line with habitat-suitability and restoration studies that emphasize the need to consider forest and grassland coverage thresholds and suitability potential when planning restoration measures in erosion-prone or ecologically fragile basins [56].

From an integrated perspective, these findings indicate that resilience in Guangdong’s cities is shaped by the interaction between a spatially uneven R–A–R baseline, nonlinear BD thresholds, and the dominant influence of forest cover and protected areas. The contraction and spatial concentration of high resistance and recoverability, together with the emergence of BD thresholds at higher richness levels, suggest that many urban and agricultural ecosystems may be operating below the levels of habitat integrity and biodiversity needed to sustain robust resilience. In such systems, incremental increases in BD, especially where habitats are fragmented or heavily modified, may not be sufficient to produce substantial gains in ER unless accompanied by improvements in forest cover, connectivity, and overall habitat quality. Conversely, in cities and regions that already possess relatively extensive and connected forest and reserve networks, BD can play a more meaningful buffering role, although the empirical effect sizes observed in this study remain modest. Similar interactions between resilience, habitat structure, and land-use intensity have been reported in studies of ecosystem climate resilience and ecosystem services, where resilience thresholds and trade-offs emerge along gradients of climate risk and land-use intensification [51,57].

These city-level patterns also have practical and policy implications. The dominance of forest coverage and nature-reserve area in explaining ER suggests that resilience-oriented management in rapidly urbanizing regions should prioritize the maintenance and enhancement of forested and protected lands, particularly where they support landscape connectivity across otherwise fragmented urban matrices. This aligns with recent work showing that integrating habitat risk and landscape resilience can be used to identify priority areas and flexible management actions for forest protection and restoration [52], and with studies that emphasize the role of green infrastructure connectivity in enhancing urban resilience through nature-based solutions [54]. The presence of BD thresholds further implies that biodiversity policies should distinguish between systems that are far below, near, or above critical richness levels. In low-BD, heavily modified landscapes, substantial habitat restoration, afforestation, and structural interventions may be necessary before BD can contribute meaningfully to resilience. In systems closer to or above the estimated thresholds, smaller-scale, targeted BD enhancement and habitat-quality improvements may help consolidate or modestly increase ER, especially where forest and reserve networks are already in place. In the context of Guangdong, these findings provide quantitative support for the province’s ecological conservation redline (ECR) policy. Empirical evaluations of ECR implementation in the Pearl River Delta indicate that redline zones effectively maintain and enhance ecosystem structural stability and functional security, with ecological quality and functions inside the redline clearly outperforming those in surrounding areas [58]. City-level assessments in Shenzhen likewise show that long-term control of land conversion within ecological control lines and redline boundaries, combined with ecological restoration projects and strengthened monitoring, has contributed to gradual improvements in ecosystem regulation services and overall ecological condition in a highly urbanized setting [59]. Against this backdrop, the present results—identifying forest-dominated, protected areas as resilience strongholds and documenting where ER has declined most sharply—offer a complementary, resilience-focused basis for refining ECR zoning and management priorities at the city scale. At the same time, the limited role of economic and short-term climatic variables in the models underscores that, within the period and region considered here, investments in habitat integrity and configuration are more directly linked to spatial differences in ER than aggregate economic growth or interannual climate variability. These conclusions are broadly consistent with regional studies that link ecological resilience decline to spatial structural conflicts and emphasize the importance of spatial control, connectivity building, and multifunctional land-use planning in coastal and rapidly urbanizing areas [47,48].

From a methodological perspective, the integration of panel threshold modeling with interpretable machine learning (XGBoost and SHAP), combined with effort normalization for observation-based BD data, contributes to ongoing efforts to capture nonlinear dynamics, thresholds, and dominant drivers in complex socio-ecological systems. Panel threshold models have been widely applied to explore nonlinear relationships between urbanization, environmental governance, and ecological resilience [60], but have rarely been used to examine ER–BD linkages at the city scale. Likewise, recent studies have employed machine learning and SHAP analysis to explore nonlinear effects and “sweet spots” in urban form–perception relationships and in climate-adaptation ecological restoration zoning [57,61], yet few have explicitly combined threshold regression with interpretable machine learning and biodiversity effort normalization. In this regard, the approach adopted here offers a complementary pathway for urban resilience research, particularly in settings where sample sizes are modest, data are heterogeneous, and both nonlinear effects and driver rankings are of interest.

Despite these contributions, the analysis remains subject to several limitations. The use of three discrete time points and city-level aggregated data may conceal finer-scale ecological dynamics, seasonal variability, and local stressors. Key factors such as long-term climate trends, detailed land-use trajectories, governance arrangements, and socio-economic inequalities are not explicitly modeled, which may affect the estimated relationships between ER, BD, and other drivers. The BD metric is derived from human observation records (e.g., GBIF), and is therefore susceptible to spatial and temporal biases associated with research capacity, economic development, and population density. In 2010 and 2015, observation intensity was especially low in some less-developed cities, likely depressing recorded species richness and limiting the ability to detect thresholds or nonlinearity. Even by 2020, observation-based BD remains uneven across space. These challenges echo more general findings that biodiversity monitoring data, including GBIF-mediated records and citizen-science platforms, exhibit strong spatial and taxonomic biases that can influence inference about biodiversity–environment relationships and temporal trends [62,63]. Future research should address these issues by integrating high-resolution spatio-temporal data, multi-dimensional BD indicators (e.g., functional and phylogenetic diversity), explicit measures of habitat connectivity, and richer socio-environmental covariates. Such extensions would help to more fully link the mechanisms highlighted in the conceptual and empirical literature on ER and BD with observed resilience dynamics in rapidly urbanizing regions and to identify critical leverage points for targeted ecological interventions.

6. Conclusions

This study provides a comprehensive analysis of the nonlinear relationship between ER and BD at the prefecture-city level in Guangdong, integrating multidimensional resilience evaluation, panel threshold modeling, and machine learning interpretation. Across 2010–2020, ER, measured through Resistance, Adaptability, and Recoverability, exhibited pronounced spatio-temporal differentiation. High-resistance, high-adaptability, and high-recoverability areas declined over time, becoming increasingly concentrated in high-elevation, forest-dominated, and less-disturbed regions. These trends reflect both the pressures of urbanization and the spatial constraints of ecosystem functionality, emphasizing the need for targeted ecological management.

The threshold analysis demonstrates that BD enhances ER in a nonlinear manner. Specifically, resilience improvements occur primarily when BD surpasses critical thresholds, which increased over the study period. This finding highlights that ecosystems require a minimum level of species richness before significant gains in resistance, adaptability, and recoverability can be realized. Incremental increases in BD below or far from these thresholds may therefore have limited ecological impact, implying that conservation efforts should be selective and context-sensitive rather than assuming linear returns to biodiversity enhancement.

Complementing these insights, the machine learning analysis identifies forest coverage ratio and the area of nature reserves as the most influential driver of ER, with BD and urban built-up area ratio also contributing significantly. This emphasizes that natural land cover and species richness act synergistically to support ecosystem stability and recovery, reinforcing the central importance of forest structure, protection status, and connectivity in resilience-oriented planning.

Overall, the study addresses a key gap in quantitative, city-level research on nonlinear ecological dynamics, offering empirical evidence of threshold effects and the prominent role of forested and protected landscapes in shaping ER relative to other drivers. The findings provide actionable guidance for policymakers and urban planners: prioritizing forest conservation, maintaining and enhancing nature reserves, improving habitat connectivity, and carefully targeted BD enhancement in systems near or below critical biodiversity levels are likely to be more effective in strengthening resilience than diffuse, small-scale interventions in already moderately biodiverse or heavily urbanized areas. At the same time, the relatively small effect sizes associated with BD in the later years call for realistic expectations about what biodiversity gains alone can achieve in the absence of broader habitat protection and restoration.

Limitations of the study include reliance on three discrete time points, aggregated city-level data, and the exclusion of additional environmental or anthropogenic factors, such as climate variability and land-use changes. Future research should incorporate higher-resolution spatial-temporal data, multi-dimensional BD indicators, and broader ecological and socio-economic drivers to refine understanding of resilience dynamics and improve evidence-based ecosystem management strategies.

In conclusion, this study demonstrates that achieving sustainable ecological resilience in rapidly urbanizing regions requires a focus on forest conservation and restoration, the protection and strategic expansion of nature reserves, and urban planning policies that maintain and enhance critical habitat structure and connectivity, complemented by targeted biodiversity conservation where it is most likely to be effective. The integrated methodological approach employed here provides a robust framework for future research and practical applications in ecological and urban resilience planning.